def call(f, *args, **kwds):
reset()
kwd_intfs, kwd_params = extract_arg_kwds(kwds, f)
args_comb = combine_arg_kwds(args, kwd_intfs, f)
paramspec = inspect.getfullargspec(f.func)
args, annotations = resolve_args(args_comb, paramspec.args,
paramspec.annotations, paramspec.varargs)
dtypes = [infer_dtype(args[arg], annotations[arg]) for arg in args]
seqs = [drv(t=t, seq=[v]) for t, v in zip(dtypes, args_comb)]
outputs = f(*seqs, **kwd_params)
if isinstance(outputs, tuple):
res = [[] for _ in outputs]
for o, r in zip(outputs, res):
collect(o | mon, result=r)
else:
res = [[]]
collect(outputs | mon, result=res[0])
sim(check_activity=False)
if isinstance(outputs, tuple):
return res
else:
return res[0]
def run_matrix(impl, mat1, mat2, cols_per_row, col_only: bool = False):
reg['trace/level'] = 0
reg['gear/memoize'] = False
# Add one more dimension to the matrix to support input type for design
mat1 = mat1.reshape(1, mat1.shape[0], mat1.shape[1])
mat2 = mat2.reshape(mat2.shape[0], 1, mat2.shape[1])
# configuration driving
cfg = create_valid_cfg(cols_per_row, mat1)
cfg_drv = drv(t=TCfg, seq=[cfg])
row_t = Queue[Array[Int[16], cfg['num_cols']]]
mat1_drv = drv(t=Queue[row_t], seq=[mat1])
res_list = []
if col_only:
# remove the extra dimension that was previously added since colum mult accepts
mat2 = np.squeeze(mat2)
# for columtn multiplication second operand needs to be only one row
mat2_drv = drv(t=row_t, seq=[mat2])
res = column_multiplication(cfg_drv, mat1_drv, mat2_drv)
# column multiplication returns result in a queue so flatening makes it a regular list
collect(res | flatten, result=res_list)
if impl == 'hw':
cosim('/column_multiplication',
'verilator',
outdir='/tmp/column_multiplication',
rebuild=True,
timeout=100)
else:
mat2_drv = drv(t=Queue[row_t], seq=[mat2])
res = matrix_multiplication(cfg_drv,
mat1_drv,
mat2_drv,
cols_per_row=cols_per_row)
collect(res, result=res_list)
if impl == 'hw':
cosim('/matrix_multiplication',
'verilator',
outdir='/tmp/matrix_multiplication',
rebuild=True,
timeout=100)
try:
sim()
# convert PG results into regular 'int'
if col_only:
pg_res = [int(el) for el in res_list]
else:
pg_res = [int(el) for row_chunk in res_list for el in row_chunk]
# calculate reference NumPy resutls
np_res = np.dot(np.squeeze(mat1), np.transpose(mat2.squeeze()))
# reshape PG results into the same format as
pg_res = np.array(pg_res).reshape(np_res.shape)
sim_assert(
np.equal(pg_res, np_res).all(), "Error in compatring results")
log.info("\033[92m //==== PASS ====// \033[90m")
except:
# printing stack trace
traceback.print_exc()
log.info("\033[91m //==== FAILED ====// \033[90m")
def test_start_stop_step_combined(sim_cls):
res = []
qrange((0, 8, 1), sim_cls=sim_cls) | collect(result=res)
sim(timeout=16)
assert res == [(i, i == 7) for i in range(8)] * 2
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
@gear
def darken(din: Uint[8], *, gain) -> Uint[8]:
return din * Ufixp[0, 8](gain)
orig_img = (mpimg.imread('../creature.png') * 255).astype(np.uint8)
res = []
drv(t=Uint[8], seq=orig_img.flatten()) \
| darken(gain=0.8) \
| int \
| collect(result=res)
reg['trace/level'] = 0
sim()
res_img = np.array(res, np.uint8)
res_img.shape = orig_img.shape
ax1 = plt.subplot(1, 2, 1)
ax1.imshow(orig_img)
ax2 = plt.subplot(1, 2, 2)
ax2.imshow(res_img)
plt.show()
def matrix_ops_single():
########################## DESIGN CONTROLS ##########################
num_cols = 8
num_rows = 6 # HINT suppoerted all dimesitions > 1
cols_per_row = 2 # HINT suported values that are divisible with num_colls
########################### TEST CONTROLS ###########################
sv_gen = 1
###########################################################################
# set either random or custom seed
seed = random.randrange(0, 2**32, 1)
# seed = 1379896999
# """Unify all seeds"""
log.info(f"Random SEED: {seed}")
set_seed(seed)
## input randomization
mat1 = np.random.randint(128, size=(num_rows, num_cols))
mat2 = np.random.randint(128, size=(num_rows, num_cols))
mat1 = np.ones((num_rows, num_cols))
mat2 = np.ones((num_rows, num_cols))
# input the constatn value optionally
# mat1 = np.empty((num_rows, num_cols))
# mat2 = np.empty((num_rows, num_cols))
# # fill the matrix with the same value
# mat1.fill(32767)
# mat2.fill(-32768)
print("Inputs: ")
print(type(mat1))
print(mat1)
print(type(mat2))
print(mat2)
reg['trace/level'] = 0
reg['gear/memoize'] = False
reg['debug/trace'] = ['*']
reg['debug/webviewer'] = True
res_list = []
cfg = {
"cols_per_row": cols_per_row,
"num_rows": num_rows,
"num_cols": num_cols,
'cols_per_multiplier': num_rows // cols_per_row
}
cfg_seq = [cfg]
cfg_drv = drv(t=TCfg, seq=cfg_seq)
# Add one more dimenstion to the matrix to support input type for design
mat1 = mat1.reshape(1, mat1.shape[0], mat1.shape[1])
mat2 = mat2.reshape(mat2.shape[0], 1, mat2.shape[1])
mat1_seq = [mat1]
mat2_seq = [mat2]
row_t = Queue[Array[Int[16], cfg['num_cols']]]
mat1_drv = drv(t=Queue[row_t], seq=mat1_seq)
mat2_drv = drv(t=Queue[row_t], seq=mat2_seq)
res = matrix_multiplication(cfg_drv,
mat1_drv,
mat2_drv,
cols_per_row=cols_per_row)
collect(res, result=res_list)
if sv_gen:
cosim('/matrix_multiplication',
'verilator',
outdir='build/matrix_multiplication/rtl',
rebuild=True,
timeout=100)
sim('build/matrix_multiplication')
## Print raw results results
log.info(f'len_res_list: \n{len(res_list)}')
try:
pg_res = [int(el) for row_chunk in res_list for el in row_chunk]
# calc refference data - matrix2 needs to be transposed before doing multiplocation
np_res = np.dot(np.squeeze(mat1), np.transpose(mat2.squeeze()))
pg_res = np.array(pg_res).reshape(np_res.shape)
log.info(f'result: \n{res}')
log.info(f'pg_res: \n{pg_res}, shape: {pg_res.shape}')
log.info(f'np_res: \n{np_res}, shape: {np_res.shape}')
sim_assert(
np.equal(pg_res, np_res).all(), "Error in compatring results")
log.info("\033[92m //==== PASS ====// \033[90m")
except:
# printing stack trace
traceback.print_exc()
log.info("\033[91m //==== FAILED ====// \033[90m")
