def common_uspmv_dev(rows,cols,datas,m_sizes,k_sizes,nnz_sizes, num_runs,A_handles,B_handles,C_handles):
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
stages = int(xclbin_opts["GEMX_uspmvStages"])
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 = gemx.addDevBuf(B_handles[0],num_runs, k_sizes[0],np.float32)
B_tmp = np.zeros((num_runs, k_sizes[0]), dtype=np.float32)
test.fillMod(9, num_runs, k_sizes[0],B_tmp)
B[:]=B_tmp
C_list=[]
for i in range(len(m_sizes)):
C = gemx.addDevBuf(C_handles[i],num_runs, m_sizes[-1],np.float32)
C.fill(0)
C_list.append(C)
A = gemx.addUSpDevBuf(np.array(rows[i]).astype(np.uint16),
np.array(cols[i]).astype(np.uint16),
np.array(datas[i]),
A_handles[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.sendDevBuf(A_handles[i])
gemx.sendDevBuf(B_handles[i])
gemx.sendDevBuf(C_handles[i])
gemx.addUSPMVDevOp(A_handles[i],B_handles[i],C_handles[i],num_runs)
gemx.executeDev()
gemx.getDevBuf(C_handles[-1])
test.multiply_and_cmp_uspmv(rows,cols,datas,m_sizes,k_sizes,B,C_list[-1])
def common_spmv(row, col, data, m, k, nnz, vector_range):
if xclbin_opts["GEMX_dataType"] == "float":
data_type = np.float32
elif xclbin_opts["GEMX_dataType"] == "int32_t":
data_type = np.int32
else:
raise TypeError("type", xclbin_opts["GEMX_dataType"], "not supported")
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
min_k = ddrWidth
spmvWidth = int(xclbin_opts["GEMX_spmvWidth"])
min_m = spmvWidth * int(xclbin_opts["GEMX_spmvMacGroups"])
m = get_padded_size(m, min_m)
k = get_padded_size(k, min_k)
print("size:", m, k, "nnz:", nnz)
if data_type == np.int32:
B = np.random.randint(low=-vector_range,
high=vector_range,
size=(k, 1),
dtype=np.int32)
else:
B = np.zeros((k, 1), dtype=np.float32)
test.fillMod(B, k, vector_range)
C = np.zeros((m, 1), dtype=data_type)
A = gemx.sendSpMat(row, col, data, m, k, nnz, xclbin_opts)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(A, B, C, nnz, xclbin_opts)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)
def common_spmv(row, col, data, m, k, nnz, vector_range, dtype):
if dtype == np.int32:
B = np.random.randint(low=-vector_range,
high=vector_range,
size=(k, 1),
dtype=np.int32)
C = np.zeros((m, 1), dtype=np.int32)
A = gemx.sendSpMat(row, col, data, nnz, dtype)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(A, B, C, nnz)
gemx.execute()
gemx.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)
elif dtype == np.float32:
B = np.zeros((k, 1), dtype=np.float32)
test.fillMod(B, k, vector_range)
C = np.zeros((m, 1), dtype=np.float32)
A = gemx.sendSpMat(row, col, data, nnz, dtype)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(A, B, C, nnz)
gemx.execute()
gemx.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)
else:
raise TypeError("type", dtype, "not supported")
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 common_spmv(row, col, data, m, k, nnz, vector_range):
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
if xclbin_opts["GEMX_dataType"] == "float":
dtype = np.float32
elif xclbin_opts["GEMX_dataType"] == "int32_t":
dtype = np.int32
else:
raise TypeError("type", xclbin_opts["GEMX_dataType"], "not supported")
if dtype == np.int32:
B = np.random.randint(low=-vector_range,
high=vector_range,
size=(k, 1),
dtype=np.int32)
C = np.zeros((m, 1), dtype=np.int32)
A = gemx.sendSpMat(row, col, data, ddrWidth, dtype)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(A, B, C, nnz)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)
elif dtype == np.float32:
B = np.zeros((k, 1), dtype=np.float32)
test.fillMod(B, k, vector_range)
C = np.zeros((m, 1), dtype=np.float32)
A = gemx.sendSpMat(row, col, data, ddrWidth, dtype)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(A, B, C, nnz)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)
def common_spmv(row, col, data, m, k, nnz, vector_range):
if xclbin_opts["GEMX_dataType"] == "float":
data_type = np.float32
elif xclbin_opts["GEMX_dataType"] == "int32_t":
data_type = np.int32
else:
raise TypeError("type", xclbin_opts["GEMX_dataType"], "not supported")
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
min_k = ddrWidth
if xclbin_opts["GEMX_useURAM"] == "1":
min_nnz = ddrWidth
min_m = ddrWidth * int(xclbin_opts["GEMX_spmvUramGroups"])
else:
spmvWidth = int(xclbin_opts["GEMX_spmvWidth"])
min_nnz = spmvWidth
min_m = spmvWidth * int(xclbin_opts["GEMX_spmvMacGroups"])
while nnz % min_nnz != 0:
# pad with 0s and adjust dimensions when necessary
row = (np.append(row, 0)).astype(np.int32)
col = (np.append(col, 0)).astype(np.int32)
data = (np.append(data, 0)).astype(np.float32)
nnz = nnz + 1
m = get_padded_size(m, min_m)
k = get_padded_size(k, min_k)
print("size:", m, k, "nnz:", nnz)
if data_type == np.int32:
B = np.random.randint(low=-vector_range,
high=vector_range,
size=(k, 1),
dtype=np.int32)
else:
B = np.zeros((k, 1), dtype=np.float32)
test.fillMod(B, k, vector_range)
C = np.zeros((m, 1), dtype=data_type)
A = xfmlp.sendSpMat(row, col, data, m, k, nnz, xclbin_opts, data_type)
xfmlp.sendMat(B)
xfmlp.sendMat(C)
xfmlp.addSPMVOp(A, B, C, nnz, xclbin_opts)
xfmlp.execute()
xfmlp.clearInstrBuf()
xfmlp.getMat(C)
test.multiply_and_cmp_spmv(row, col, data, m, k, nnz, B, C)