def __init__(self, xclbin_opt, wgt, bias, wgt_scale, post_scale):
#Ensuring min_m and min_n never fall below min_k is needed when chaining multiple GEMM operations
#If min_m/min_n is less than min_k, using the output of a GEMM call where either dimension
#is less than min_k would lead to bad results if it's directly used as input for another GEMM operation
self.min_m = 32 * max(int(xclbin_opt["GEMX_gemmKBlocks"]),
int(xclbin_opt["GEMX_gemmMBlocks"]))
self.min_k = 32 * int(xclbin_opt["GEMX_gemmKBlocks"])
self.min_n = 32 * max(int(xclbin_opt["GEMX_gemmKBlocks"]),
int(xclbin_opt["GEMX_gemmNBlocks"]))
if type(wgt) != list:
wgt = [wgt]
if type(bias) != list:
bias = [bias]
self._wshape = []
for w in wgt:
self._wshape.append(w.shape)
self._qw = [np.int16(a * b) for a, b in zip(wgt, wgt_scale)]
self._qb = [np.int32(a * b) for a, b in zip(bias, wgt_scale)]
for i, b in enumerate(self._qw):
self._qw[i] = self.format_for_fpga(b, self.min_k, self.min_n)
gemx.sendMat(self._qw[i])
#in_row, in_col = self.get_padded_shape(in_dim, self.min_m, self.min_k)
self.fpga_buf = []
self.out_dim = None
self.post_scale = post_scale
self.batch_sz = 0
def format_bias(self, b, dim, min_row, min_col):
if b.ndim == 1:
b = np.broadcast_to(b, dim)
b = self.format_for_fpga(b, min_row, min_col)
gemx.sendMat(b)
return b
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):
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 format_bias(self, b, dim, min_row, min_col):
if b.ndim == 1:
b = np.broadcast_to(b, (dim[1], dim[0]))
b = np.transpose(b)
b = self.format_for_fpga(b, min_row, min_col)
gemx.sendMat(b)
return b
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 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')
gemx.sendMat(mat_A, PE)
gemx.sendMat(mat_B, PE)
gemx.sendMat(C_fpga, PE)
gemx.sendMat(bias, PE)
gemx.addFCNOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1],
RELU_scale[0], RELU_scale[1], PE)
gemx.execute(PE)
gemx.clearInstrBuf(PE)
gemx.getMat(C_fpga, PE)
self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale,
RELU_scale)
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_gemm(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)
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)
start_time = time.time()
gemx.sendMat(mat_A)
gemx.sendMat(mat_B)
gemx.sendMat(C_fpga)
gemx.sendMat(bias)
gemx.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1])
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_basic(self,
PE,
xclbin_opts,
mat_A,
mat_B,
bias,
post_scale=[1, 0]):
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))
if xclbin_opts["GEMX_dataType"] == "short":
C_fpga = np.zeros((m, n), dtype=np.int16, order='C')
else: #float
C_fpga = np.zeros((m, n), dtype=np.float32, order='C')
gemx.sendMat(mat_A, PE)
gemx.sendMat(mat_B, PE)
gemx.sendMat(C_fpga, PE)
gemx.sendMat(bias, PE)
gemx.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0],
post_scale[1],
PE) # default test_basic will call addGEMMOp
gemx.execute(PE)
gemx.clearInstrBuf(PE)
gemx.getMat(C_fpga, PE)
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):
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, 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 predict(self, inp, in_scale):
self.init_fpgabuf(inp.shape)
self.loadInstr()
padded_arr = self.format_for_fpga(inp * in_scale, self.min_m,
self.min_k)
#print ("input shape", padded_arr.shape)
np.copyto(self.fpga_buf[0],
np.int16(padded_arr),
casting='same_kind',
where=True)
gemx.sendMat(self.fpga_buf[0])
gemx.execute()
gemx.getMat(self.fpga_buf[-1])
return self.fpga_buf[-1][:self.out_dim[0], :self.out_dim[1]]
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 predict ( self, inp, in_scale, post_scale):
row_padded, col_padded = self.get_padded_shape( inp.shape, self.min_m, self.min_k)
padded_arr = np.zeros ( (row_padded, col_padded), dtype=inp.dtype, order='C')
padded_arr[0:inp.shape[0], 0:inp.shape[1]] = inp
print ("input shape", padded_arr.shape)
np.copyto(self.fpga_buf[0], np.int16( padded_arr * in_scale ), casting='same_kind', where=True)
gemx.sendMat(self.fpga_buf[0])
for i,l in enumerate(self.kmodel.layers):
act = l.get_config()['activation']
if act == 'relu':
gemx.addFCNOp( self.fpga_buf[i], self.w[i], self.fpga_buf[i+1], self.b[i], post_scale[i][0], post_scale[i][1], 0, 0)
else:
gemx.addGEMMOp( self.fpga_buf[i], self.w[i], self.fpga_buf[i+1], self.b[i], post_scale[i][0], post_scale[i][1])
gemx.execute()
gemx.getMat (self.fpga_buf[-1])
return self.fpga_buf[-1][:self.out_dim[0],:self.out_dim[1]]
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)
def __init__(self, xclbin_opts, wgt, bias, wgt_scale, bias_scale,
post_scale, relu_scale):
#Ensuring min_m and min_n never fall below min_k is needed when chaining multiple GEMM operations
#If min_m/min_n is less than min_k, using the output of a GEMM call where either dimension
#is less than min_k would lead to bad results if it's directly used as input for another GEMM operation
ddrwidth = int(xclbin_opts["GEMX_ddrWidth"])
self.min_m = ddrwidth * max(int(xclbin_opts["GEMX_gemmKBlocks"]),
int(xclbin_opts["GEMX_gemmMBlocks"]))
self.min_k = ddrwidth * int(xclbin_opts["GEMX_gemmKBlocks"])
self.min_n = ddrwidth * int(xclbin_opts["GEMX_gemmNBlocks"])
if type(wgt) != list:
wgt = [wgt]
if type(bias) != list:
bias = [bias]
assert len(wgt) == len(wgt_scale)
assert len(bias) == len(bias_scale)
self._wshape = []
for w in wgt:
self._wshape.append(w.shape)
if xclbin_opts["GEMX_dataType"] == "float":
self._qw = wgt
self._qb = bias
else:
self._qw = [
np.int16(np.around(a * b)) for a, b in zip(wgt, wgt_scale)
]
self._qb = [
np.int32(np.around(a * b)) for a, b in zip(bias, bias_scale)
]
for i, b in enumerate(self._qw):
b = np.transpose(b)
self._qw[i] = self.format_for_fpga(b, self.min_m, self.min_k)
gemx.sendMat(self._qw[i])
#in_row, in_col = self.get_padded_shape(in_dim, self.min_m, self.min_k)
self.fpga_buf = []
self.out_dim = None
self.post_scale = post_scale
self.relu_scale = relu_scale
self.batch_sz = 0
def predict(self, inp, in_scale):
"""
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
in_scale: float
input scale
Return
------
array
result prediction matrix
"""
inp = np.transpose(inp)
self.init_fpgabuf(inp.shape)
self.loadInstr()
padded_arr = self.format_for_fpga(inp * in_scale, self.min_k,
self.min_n)
#print ("input shape", padded_arr.shape)
if self.fpga_buf[0].dtype == np.int16:
np.copyto(self.fpga_buf[0],
np.int16(padded_arr),
casting='same_kind',
where=True)
else:
np.copyto(self.fpga_buf[0],
padded_arr,
casting='same_kind',
where=True)
gemx.sendMat(self.fpga_buf[0])
gemx.execute()
gemx.getMat(self.fpga_buf[-1])
return np.transpose(
self.fpga_buf[-1][:self.out_dim[0], :self.out_dim[1]])
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] n = mat_B.shape[1] C_fpga = np.zeros((m, n), dtype=np.int16, order='C') 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) self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def predict(self, inp, in_scale):
C_list = [[]] * (len(self.kmodel.layers) + 1)
inp = self.format_for_fpga(inp, self.min_m, self.min_m)
C_list[0] = np.transpose(inp * in_scale)
for i, bi in enumerate(self._qb):
bi = bi.reshape(bi.shape[0], 1)
bi = self.format_for_fpga(bi, C_list[i].shape[1], self.sizes[i][0])
for j in range(C_list[i].shape[1]):
B = (C_list[i][:, j]).astype(np.float32)
C = np.zeros((self.sizes[i][0], 1), dtype=np.float32)
gemx.sendMat(B)
gemx.sendMat(C)
gemx.addSPMVOp(self.A_list[i], B, C, self.sizes[i][2])
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C)
if j == 0:
C_list[i + 1] = C
else:
C_list[i + 1] = np.append(C_list[i + 1], C, axis=1)
C_list[i + 1] = C_list[i + 1] + np.transpose(bi)
result = np.transpose(C_list[-1])
return result[:self.out_dim[0], :self.out_dim[1]]
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)
gemx.sendMat(mat_A,PE)
gemx.sendMat(mat_B,PE)
gemx.sendMat(C_fpga,PE)
gemx.sendMat(bias, PE)
gemx.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1], PE) # default test_basic will call addGEMMOp
gemx.execute(PE)
gemx.getMat(C_fpga,PE)
self.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
gemx.sendMat(mat_A)
gemx.sendMat(mat_B)
gemx.sendMat(C_fpga)
gemx.sendMat(bias)
gemx.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1])
timePointKernel.append(time.time()) # send to FPGA
gemx.execute()
gemx.clearInstrBuf()
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)
test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
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
"""
C_list = [[]] * 2
inp = self.format_for_fpga(inp, 1, 1)
C_list[0] = np.transpose(inp)
B = (C_list[0][:, 0]).astype(np.float32)
C_vector = [B]
gemx.sendMat(C_vector[0])
for i, l in enumerate(self.kmodel.layers):
C_vector.append(np.zeros((self.sizes[i][0], 1), dtype=np.float32))
gemx.sendMat(C_vector[i + 1])
activation = True if l.get_config(
)['activation'] == 'relu' else False
gemx.addSPMVOp(self.A_list[i], C_vector[i], C_vector[i + 1],
self.sizes[i][2], self.xclbin_opts, activation)
gemx.execute()
gemx.getMat(C_vector[-1])
C_list[1] = C_vector[-1]
for j in range(1, C_list[0].shape[1]):
C_vector[0][:] = (C_list[0][:, j]).astype(np.float32)
gemx.sendMat(C_vector[0])
C_vector[-1].fill(0)
for i in range(len(self.kmodel.layers)):
gemx.sendMat(C_vector[i + 1])
gemx.execute()
gemx.getMat(C_vector[-1])
C_list[1] = np.append(C_list[1], C_vector[-1], axis=1)
gemx.clearInstrBuf()
result = np.transpose(C_list[1])
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]]
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)
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)
gemx.sendMat(A)
gemx.sendMat(B)
gemx.sendMat(b0)
gemx.sendMat(C)
gemx.sendMat(D)
gemx.sendMat(E)
gemx.sendMat(b1)
gemx.addGEMMOp(A, B, C, b0, 1, 0)
gemx.addGEMMOp(D, C, E, b1, 1, 0)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C)
gemx.getMat(E)
print("test C")
test.multiply_and_cmp(C, A, B, b0, m, n, [1, 0])
print("test E")
test.multiply_and_cmp(E, D, C, b1, m, n, [1, 0])
int(xclbin_opt["GEMX_gemmKBlocks"]), int(
xclbin_opt["GEMX_gemmNBlocks"]))
if args.mtx == 'none':
num_matrix = len(args.matrix) / 3
for i in range(num_matrix):
m = args.matrix[i * 3]
k = args.matrix[i * 3 + 1]
n = args.matrix[i * 3 + 2]
m = int(math.ceil(np.float32(m) / min_m) * min_m)
k = int(math.ceil(np.float32(k) / min_k) * min_k)
n = int(math.ceil(np.float32(n) / min_n) * min_n)
print(m, k, n)
A = np.zeros((m, k), dtype=np.int16)
A.fill(1)
A_buf.append(A)
gemx.sendMat(A_buf[i])
B_buf.append(np.zeros((k, n), dtype=np.int16))
C_buf.append(np.zeros((m, n), dtype=np.int16))
bias = np.zeros((m, n), dtype=np.int32)
bias.fill(1)
bias_buf.append(bias)
else:
#For fcn if read from mtx files, matrix sizes still need to be provided
num_matrix = len(args.mtx)
if num_matrix != len(args.matrix) / 3:
raise Exception("please enter sizes for each layer")
for i in range(num_matrix):
matA = sio.mmread(args.mtx[i])
m = args.matrix[i * 3]
k = args.matrix[i * 3 + 1]
n = args.matrix[i * 3 + 2]
