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_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 test_perf_fcn(m,
k,
n,
xclbin_opts,
post_scale=[1, 0],
A_range=32764,
B_range=32764,
bias_range=32764):
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
m = test.get_padded_size(m,
int(xclbin_opts["GEMX_gemmMBlocks"]) * ddrWidth)
k = test.get_padded_size(k,
int(xclbin_opts["GEMX_gemmKBlocks"]) * ddrWidth)
n = test.get_padded_size(n,
int(xclbin_opts["GEMX_gemmNBlocks"]) * ddrWidth)
if xclbin_opts["GEMX_dataType"] == "short":
mat_A = np.random.randint(low=-A_range,
high=A_range,
size=(m, k),
dtype=np.int16)
mat_B = np.random.randint(low=-B_range,
high=B_range,
size=(k, n),
dtype=np.int16)
bias = []
if bias_range != 0:
bias = np.random.randint(low=-bias_range,
high=bias_range,
size=(m, n),
dtype=np.int32)
else:
bias = np.zeros((m, n), dtype=np.int32, order='C')
C_fpga = np.zeros((m, n), dtype=np.int16)
else:
mat_A = np.random.uniform(low=-128, high=128,
size=(m, k)).astype(np.float32)
mat_B = np.random.uniform(low=-128, high=128,
size=(k, n)).astype(np.float32)
bias = np.zeros((m, n), dtype=np.float32, order='C')
C_fpga = np.zeros((m, n), dtype=np.float32)
start_time = time.time()
gemx.sendMat(mat_A)
gemx.sendMat(mat_B)
gemx.sendMat(C_fpga)
gemx.sendMat(bias)
gemx.addFCNOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1], 1,
0)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C_fpga)
end_time = time.time()
total_operations = 2 * m * n * k + m * n * 3
test.test_perf(end_time - start_time, total_operations, m, k, n, ddrWidth)
test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_multi_fcn(ins_count,
m_size,
k_size,
n_size,
post_scale=[1, 0],
A_range=32764,
B_range=32764):
mat_A = []
mat_C = []
mat_bias = []
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
for i in range(ins_count):
m_size[i] = test.get_padded_size(
m_size[i],
int(xclbin_opts["GEMX_gemmMBlocks"]) * ddrWidth)
k_size[i] = test.get_padded_size(
k_size[i],
int(xclbin_opts["GEMX_gemmKBlocks"]) * ddrWidth)
n_size[i] = test.get_padded_size(
n_size[i],
int(xclbin_opts["GEMX_gemmNBlocks"]) * ddrWidth)
mat_A.append(
np.random.randint(low=-A_range,
high=A_range,
size=(m_size[i], k_size[i]),
dtype=np.int16))
mat_bias.append(np.zeros((m_size[i], n_size[i]), dtype=np.int32))
mat_C.append(
np.zeros((m_size[i], n_size[i]), dtype=np.int16, order='C'))
mat_B0 = np.random.randint(low=-B_range,
high=B_range,
size=(k_size[0], n_size[0]),
dtype=np.int16)
for i in range(ins_count):
gemx.sendMat(mat_A[i])
gemx.sendMat(mat_C[i])
gemx.sendMat(mat_bias[i])
gemx.sendMat(mat_B0)
gemx.addFCNOp(mat_A[0], mat_B0, mat_C[0], mat_bias[0], post_scale[0],
post_scale[1], 1, 0)
gemx.addFCNOp(mat_A[1], mat_C[0], mat_C[1], mat_bias[1], post_scale[0],
post_scale[1], 1, 0)
gemx.addFCNOp(mat_A[2], mat_C[1], mat_C[2], mat_bias[2], post_scale[0],
post_scale[1], 1, 0)
gemx.addFCNOp(mat_A[3], mat_C[2], mat_C[3], mat_bias[3], post_scale[0],
post_scale[1], 1, 0)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(mat_C[0])
gemx.getMat(mat_C[1])
gemx.getMat(mat_C[2])
gemx.getMat(mat_C[3])
test.multiply_and_cmp(mat_C[3], mat_A[3], mat_C[2], mat_bias[3], m_size[3],
n_size[3], post_scale)