def test_perf_fcn(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
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)
timePointKernel.append(time.time()) # send to FPGA
gemx.execute()
timePointKernel.append(time.time()) # call kernel
gemx.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 = gemx.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_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):
gemx.sendMat(mat_A[i])
gemx.sendMat(mat_C[i])
gemx.sendMat(mat_bias[i])
gemx.sendMat(mat_B0)
gemx.addGEMMOp(mat_A[0], mat_B0, mat_C[0], mat_bias[0], post_scale[0],
post_scale[1])
gemx.addGEMMOp(mat_A[1], mat_C[0], mat_C[1], mat_bias[1], post_scale[0],
post_scale[1])
gemx.addGEMMOp(mat_A[2], mat_C[1], mat_C[2], mat_bias[2], post_scale[0],
post_scale[1])
gemx.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
gemx.execute()
timePointKernel.append(time.time()) # call kernel
gemx.getMat(mat_C[0])
gemx.getMat(mat_C[1])
gemx.getMat(mat_C[2])
gemx.getMat(mat_C[3])
timePointKernel.append(time.time()) # copy from FPGA
freq = gemx.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)
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_perf_multi_fcn(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):
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)
timePointKernel.append(time.time()) # send to FPGA
gemx.execute()
gemx.clearInstrBuf()
timePointKernel.append(time.time()) # call kernel
gemx.getMat(mat_C[0])
gemx.getMat(mat_C[1])
gemx.getMat(mat_C[2])
gemx.getMat(mat_C[3])
timePointKernel.append(time.time()) # copy from FPGA
freq = gemx.getFreq()
test.test_perf(timePointKernel,total_operations,total_parallel_operations,freq,0,0,0)
test.multiply_and_cmp(mat_C[3], mat_A[3], mat_C[2], mat_bias[3], m_size[3], n_size[3], post_scale)