print('a * b =')
print(mul_nd.asnumpy())
np.testing.assert_allclose(mul_nd.asnumpy(), a_np * b_np.astype(dtype_w))
print('a / b =')
print(div_nd.asnumpy())
# numpy always round down, while in c, the numerator will be rounded to zero.
#np.testing.assert_allclose(div_nd.asnumpy(), a_np / b_np)
print('b / a =')
print(rdiv_nd.asnumpy())
print('max(a, b)=')
print(gtm_nd.asnumpy())
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description='test of NNPU Op')
parser.add_argument('--sim',
type=str,
help='the simulator to use',
default='S0',
choices=['S0', 'S1', 'SC'])
args = parser.parse_args()
with nnpu.Environment('./nnpu_config.yaml'):
env = nnpu.get_env()
nnpu.set_device(env, type=args.sim)
test()
def max_pooling(inshape, outshape, cell_shape, innp, outdetype):
ret = np.full(outshape, np.iinfo(outdetype).min, dtype=outdetype)
for w in range(outshape[0]):
for h in range(outshape[1]):
for j in range(cell_shape):
for k in range(cell_shape):
for l in range(outshape[2]):
ret[w][h][l] = max(
ret[w][h][l],
innp[w * cell_shape + j][h * cell_shape + k][l])
return ret
# reduce max
with nnpu.Environment(cfg_path):
env = nnpu.get_env()
nnpu.set_device(env, type=args.sim)
nnpu.set_dump(False)
#==================================#
# ------ first define shapes ------
#==================================#
# input data layout: HWC
in_shape = (40, 40, 256)
# pooling windows size, height == width.
cell_shape = 2
# in this demo we don't do padding, so input data height and width must be divisible to pooling window size.
assert in_shape[0] % cell_shape == 0, 'error'
assert in_shape[1] % cell_shape == 0, 'error'
help='enable profiling',
default=True)
parser.add_argument('--parallel',
type=bool,
help='enable parallel',
default=False)
parser.add_argument('--dimx', type=int, help='tile size of x', default=2)
parser.add_argument('--dimy', type=int, help='tile size of y', default=2)
args = parser.parse_args()
cfg_path = './nnpu_config-opt.yaml'
if (args.profile):
profile_dir = '/home/jian/Documents/nnpu_profile'
nnpu.set_profile(['timeline', 'memory_access_latency'], profile_dir)
with ScheduleProcHelper(), nnpu.Environment(cfg_path):
env = nnpu.get_env()
nnpu.set_device(env, type=args.sim)
shape1 = (128, 256)
shape2 = (128, 256)
macops = shape1[0] * shape1[1] * shape2[0]
gemm_shape = (8, 8, 8)
factor = gemm_shape[1]
assert shape1[1] == shape2[1], \
'gemm do dot product between rows, so the shape[1] of inputs should match'
assert shape1[0] % gemm_shape[
0] == 0, 'gemm insn require size of input 1 be x{0}'.format(
gemm_shape[0])
assert shape2[0] % gemm_shape[
2] == 0, 'gemm insn require size of input 2 be x{0}'.format(
import nnpu
import tvm
import topi
from nnpu.utils import ScheduleProcHelper
import numpy as np
import math
import argparse
parser = argparse.ArgumentParser(description='test gemm with tiled/non-tiled data')
parser.add_argument('--sim', type=str, help='the simulator to use',
default='S0', choices=['S0', 'S1', 'SC'])
args = parser.parse_args()
with ScheduleProcHelper(), nnpu.Environment('./nnpu_config_fp32.yaml'):
env = nnpu.get_env()
nnpu.set_device(env, type=args.sim)
dtype_n, dtype_w = env.cfg['dtype_n'], env.cfg['dtype_w']
assert dtype_w in ['float32', 'float16'], 'when testing activation function, float dtype is needed'
shape = (64, )
a = tvm.placeholder(shape, dtype_w, 'a')
a_buf = tvm.compute(shape, lambda *i: a(*i), 'a_buf')
exp = tvm.compute(shape, lambda i: tvm.exp(a_buf[i]), 'exp')
log = tvm.compute(shape, lambda i: tvm.log(a_buf[i]), 'exp')
tanh = tvm.compute(shape, lambda i: tvm.tanh(a_buf[i]), 'exp')
sigmoid = tvm.compute(shape, lambda i: tvm.sigmoid(a_buf[i]), 'exp')
# k = tvm.reduce_axis((0, 16), 'k0')
# sum = tvm.compute((1, ), lambda i: tvm.sum(sigmoid[k], axis=k), 'sum')
parser.add_argument('--profile',
type=bool,
help='enable profiling',
default=True)
args = parser.parse_args()
if (args.profile):
profile_dir = '/home/jian/Documents/nnpu_profile'
nnpu.set_profile(['timeline', 'memory_access_latency'], profile_dir)
def roundup(x, d):
return (x + d - 1) // d * d
with ScheduleProcHelper(), nnpu.Environment('./nnpu_config-dense.yaml'):
env = nnpu.get_env()
nnpu.set_device(env, type=args.sim)
shape1 = (128, 512)
shape2 = (129, 512)
macops = shape1[0] * shape1[1] * shape2[0]
# upround shapes
gemm_shape = (8, 8, 8)
# shape1 = (roundup(shape1[0], gemm_shape[0]), shape1[1])
shape2 = (roundup(shape2[0], gemm_shape[2]), shape2[1])
bias_shape = (shape2[0], )
factor = gemm_shape[1]
