def test_basic(self,
PE,
mat_A,
mat_B,
bias,
post_scale=[1, 1],
RELU_scale=[1, 0]):
m = mat_A.shape[0]
k = mat_A.shape[1]
n = mat_B.shape[1]
print("test Fcn")
print("test_basic: %d %d %d %d %d" %
(m, k, n, post_scale[0], post_scale[1]))
print("A: ", np.amax(mat_A), np.amin(mat_A), np.average(mat_A))
print("B: ", np.amax(mat_B), np.amin(mat_B), np.average(mat_B))
print("bias: ", np.amax(bias), np.amin(bias), np.average(bias))
C_fpga = np.zeros((m, n), dtype=np.int16, order='C')
xfmlp.sendMat(mat_A, PE)
xfmlp.sendMat(mat_B, PE)
xfmlp.sendMat(C_fpga, PE)
xfmlp.sendMat(bias, PE)
xfmlp.addFCNOp(mat_A, mat_B, C_fpga, bias, post_scale[0],
post_scale[1], RELU_scale[0], RELU_scale[1], PE)
xfmlp.execute(PE)
xfmlp.clearInstrBuf(PE)
xfmlp.getMat(C_fpga, PE)
self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale,
RELU_scale)
def test_multiInstrv1(int_range, m, k, n, add_bias=False):
print ("test_multiInstrv1: %d %d %d %d" % (int_range, m, k, n))
A = np.random.randint(low=-int_range, high=int_range, size=(m, k), dtype=np.int16)
B = np.random.randint(low=-int_range, high=int_range, size=(k, n), dtype=np.int16)
C = np.zeros ((m, n), dtype=np.int16);
D = np.random.randint(low=-int_range, high=int_range, size=(m, k), dtype=np.int16)
E = np.zeros ((m, n), dtype=np.int16);
b0 = np.zeros ((m, n), dtype=np.int32);
b1 = np.zeros ((m, n), dtype=np.int32);
if add_bias == True:
b0 = np.random.randint(low=-int_range, high=int_range, size=(m, n), dtype=np.int32)
b1 = np.random.randint(low=-int_range, high=int_range, size=(m, n), dtype=np.int32)
xfmlp.sendMat(A)
xfmlp.sendMat(B)
xfmlp.sendMat(b0)
xfmlp.sendMat(C)
xfmlp.sendMat(D)
xfmlp.sendMat(E)
xfmlp.sendMat(b1)
xfmlp.addFCNOp(A, B, C, b0, 1, 13, 307, 10)
xfmlp.addFCNOp(D, C, E, b1, 1, 18, 307, 10)
xfmlp.execute()
xfmlp.clearInstrBuf()
xfmlp.getMat(C)
xfmlp.getMat(E)
print("test C")
test.multiply_and_cmp(C, A, B, b0, m, n, [1, 13],[307, 10])
print("test E")
test.multiply_and_cmp(E, D, C, b1, m, n, [1, 18],[307, 10])
def test_perf_gemm_gemm(A_range, B_range, bias_range, m, k, n, post_scale):
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)
timePointKernel = []
timePointKernel.append(time.time()) # current time
xfmlp.sendMat(mat_A)
xfmlp.sendMat(mat_B)
xfmlp.sendMat(C_fpga)
xfmlp.sendMat(bias)
xfmlp.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1])
timePointKernel.append(time.time()) # send to FPGA
xfmlp.execute()
timePointKernel.append(time.time()) # call kernel
xfmlp.getMat(C_fpga)
timePointKernel.append(time.time()) # copy from FPGA
total_operations = 2 * m * n * k + m * n * 3
total_parallel_operations = 2 * m * n * k
freq = xfmlp.getFreq()
test.test_perf(timePointKernel,total_operations,total_parallel_operations,freq,m,k,n)
if m > 4096 and n > 4096 and k > 4096:
print("Skip golden comparision because large matrix size")
else:
test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_perf_gemm(m, k, n, A_range=32764, B_range=32764, bias_range=32764, post_scale=[1,0]):
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)
timePointKernel = []
timePointKernel.append(time.time()) # current time
xfmlp.sendMat(mat_A)
xfmlp.sendMat(mat_B)
xfmlp.sendMat(C_fpga)
xfmlp.sendMat(bias)
xfmlp.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1])
timePointKernel.append(time.time()) # send to FPGA
xfmlp.execute()
xfmlp.clearInstrBuf()
timePointKernel.append(time.time()) # call kernel
xfmlp.getMat(C_fpga)
timePointKernel.append(time.time()) # copy from FPGA
total_operations = 2 * m * n * k + m * n * 3
total_parallel_operations = 2 * m * n * k
freq = xfmlp.getFreq()
test.test_perf(timePointKernel,total_operations,total_parallel_operations,freq,m,k,n)
test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_textfiles(self, path_to_a, path_to_b, path_to_bias, post_scale):
mat_A = np.loadtxt(path_to_a, dtype=np.int16)
mat_B = np.loadtxt(path_to_b, dtype=np.int16)
bias = np.loadtxt(path_to_bias, dtype=np.int32)
m = mat_A.shape[0]
k = mat_A.shape[1]
n = mat_B.shape[1]
C_fpga = np.zeros((m, n), dtype=np.int16, order='C')
xfmlp.sendMat(mat_A)
xfmlp.sendMat(mat_B)
xfmlp.sendMat(C_fpga)
xfmlp.sendMat(bias)
xfmlp.addFCNOp(mat_A, mat_B, C_fpga, bias, post_scale[0],
post_scale[1], 1, 0)
xfmlp.execute()
xfmlp.clearInstrBuf()
xfmlp.getMat(C_fpga)
self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
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)
def test_basic(self, PE, mat_A, mat_B, bias, post_scale=[1, 1]):
m = mat_A.shape[0]
k = mat_A.shape[1]
n = mat_B.shape[1]
print("test_basic(PE=%d): %d %d %d %d %d" %
(PE, m, k, n, post_scale[0], post_scale[1]))
print("A: ", np.amax(mat_A), np.amin(mat_A), np.average(mat_A))
print("B: ", np.amax(mat_B), np.amin(mat_B), np.average(mat_B))
print("bias: ", np.amax(bias), np.amin(bias), np.average(bias))
C_fpga = np.zeros((m, n), dtype=np.int16)
xfmlp.sendMat(mat_A, PE)
xfmlp.sendMat(mat_B, PE)
xfmlp.sendMat(C_fpga, PE)
xfmlp.sendMat(bias, PE)
xfmlp.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0],
post_scale[1],
PE) # default test_basic will call addGEMMOp
xfmlp.execute(PE)
xfmlp.clearInstrBuf(PE)
xfmlp.getMat(C_fpga, PE)
self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_perf_multi_gemm(ins_count, m_size, k_size, n_size, A_range, B_range, post_scale):
total_operations = 0
total_parallel_operations = 0
mat_A=[]
mat_C=[]
mat_bias=[]
for i in range(ins_count):
total_operations += 2 * m_size[i] * n_size[i] * k_size[i] + m_size[i] * n_size[i] * 3
total_parallel_operations += 2 * m_size[i] * n_size[i] * k_size[i]
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)
timePointKernel = []
timePointKernel.append(time.time()) # current time
for i in range(ins_count):
xfmlp.sendMat(mat_A[i])
xfmlp.sendMat(mat_C[i])
xfmlp.sendMat(mat_bias[i])
xfmlp.sendMat(mat_B0)
xfmlp.addGEMMOp (mat_A[0], mat_B0, mat_C[0], mat_bias[0], post_scale[0], post_scale[1])
xfmlp.addGEMMOp (mat_A[1], mat_C[0], mat_C[1], mat_bias[1], post_scale[0], post_scale[1])
xfmlp.addGEMMOp (mat_A[2], mat_C[1], mat_C[2], mat_bias[2], post_scale[0], post_scale[1])
xfmlp.addGEMMOp (mat_A[3], mat_C[2], mat_C[3], mat_bias[3], post_scale[0], post_scale[1])
timePointKernel.append(time.time()) # send to FPGA
xfmlp.execute()
timePointKernel.append(time.time()) # call kernel
xfmlp.getMat(mat_C[0])
xfmlp.getMat(mat_C[1])
xfmlp.getMat(mat_C[2])
xfmlp.getMat(mat_C[3])
timePointKernel.append(time.time()) # copy from FPGA
freq = xfmlp.getFreq()
test.test_perf(timePointKernel,total_operations,total_parallel_operations,freq,0,0,0)
if np.max(m_size) > 4096 and np.max(k_size) > 4096 and np.max(n_size) > 4096:
print("Skip golden comparision because large matrix size")
else:
test.multiply_and_cmp(mat_C[3], mat_A[3], mat_C[2], mat_bias[3], m_size[3], n_size[3], post_scale)
print(m, k, n)
A_buf.append(np.zeros((m, k), dtype=np.int16, order='C'))
bias_buf.append(np.zeros((m, n), dtype=np.int32, order='C'))
B_buf.append(np.zeros((k, n), dtype=np.int16, order='C'))
C_buf.append(np.zeros((m, n), dtype=np.int16, order='C'))
for i in range(num_matrix):
xfmlp.sendMat(B_buf[i])
xfmlp.sendMat(A_buf[i])
xfmlp.sendMat(C_buf[i])
xfmlp.sendMat(bias_buf[i])
time.sleep(2)
total_time = 0
for k in range(args.numiter):
start_time = time.time()
xfmlp.sendMat(B_buf[0])
for i in range(num_matrix):
#xfmlp.addFCNOp(A_buf[i], B_buf[i], B_buf[i+1], bias_buf[i], 1,0,1,0 )
xfmlp.addFCNOp(A_buf[i], B_buf[i], C_buf[i], bias_buf[i], 1, 0, 1, 0)
xfmlp.execute()
xfmlp.getMat(C_buf[num_matrix - 1])
#xfmlp.wait()
total_time += time.time() - start_time
print("Average FPGA exec time(python): ", (total_time / args.numiter) * 1000,
" ms")
xfmlp.printStats()