def test_conv2d_nchw():
schedules = [
SpatialPack(vh=7, vw=1, vc=16, ba=64, bc=4, unroll=True),
SpatialPack(vh=7, vw=1, vc=16, ba=64, bc=1, unroll=True),
SpatialPack(vh=7, vw=7, vc=16, ba=16, bc=64, unroll=False)
]
# verify_conv2d_nchw_all(1, 64, 56, 64, 3, 1, 1, schedules[0])
verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, schedules[0])
# verify_conv2d_nchw_gemm(1, 64, 56, 64, 3, 1, 1)
# ResNet18 worklaods
"""
def test_resnet_1():
print(
'MKLDNN - g1mb1_ic64ih56iw56_oc64oh56ow56_kh3kw3_sh1sw1_ph1pw1_n : 0.72514 ms'
)
sch = SpatialPack(vh=7, vw=1, vc=16, ba=64, bc=4, unroll=True)
print('Run separately ...')
verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, sch)
print('Run together ...')
verify_conv2d_nchw_all(1, 64, 56, 64, 3, 1, 1, sch)
def test_resnet_0():
print(
'MKLDNN - g1mb1_ic3ih224iw224_oc64oh112ow112_kh7kw7_sh2sw2_ph3pw3_n : 0.773982 ms'
)
sch = SpatialPack(vh=28, vw=8, vc=16, ba=224, bc=1,
unroll=True) # 1.42711 ms
print('Run separately ...')
verify_conv2d_nchw(1, 3, 224, 64, 7, 2, 3, sch)
print('Run together ...')
verify_conv2d_nchw_all(1, 3, 224, 64, 7, 2, 3, sch)
Workload('int8', 'int32', 224, 224, 3, 64, 7, 7, 3, 3, 2, 2),
Workload('int8', 'int32', 56, 56, 64, 64, 3, 3, 1, 1, 1, 1),
Workload('int8', 'int32', 56, 56, 64, 64, 1, 1, 0, 0, 1, 1),
Workload('int8', 'int32', 56, 56, 64, 128, 3, 3, 1, 1, 2, 2),
Workload('int8', 'int32', 56, 56, 64, 128, 1, 1, 0, 0, 2, 2),
Workload('int8', 'int32', 28, 28, 128, 128, 3, 3, 1, 1, 1, 1),
Workload('int8', 'int32', 28, 28, 128, 256, 3, 3, 1, 1, 2, 2),
Workload('int8', 'int32', 28, 28, 128, 256, 1, 1, 0, 0, 2, 2),
Workload('int8', 'int32', 14, 14, 256, 256, 3, 3, 1, 1, 1, 1),
Workload('int8', 'int32', 14, 14, 256, 512, 3, 3, 1, 1, 2, 2),
Workload('int8', 'int32', 14, 14, 256, 512, 1, 1, 0, 0, 2, 2),
Workload('int8', 'int32', 7, 7, 512, 512, 3, 3, 1, 1, 1, 1),
]
_SCHEDULES = [
# float32 imagenet
SpatialPack(1, 8, 4, 1, 4, True),
SpatialPack(1, 8, 4, 1, 4, True),
SpatialPack(1, 7, 4, 2, 4, True),
SpatialPack(1, 4, 8, 4, 1, True),
SpatialPack(1, 4, 4, 1, 16, False),
SpatialPack(1, 4, 8, 4, 8, False),
SpatialPack(1, 7, 4, 3, 8, True),
SpatialPack(1, 2, 8, 1, 8, True),
SpatialPack(2, 1, 16, 1, 4, True),
SpatialPack(1, 7, 4, 1, 1, True),
Im2ColPack(7, 4, 1, 16, True),
Im2ColPack(7, 4, 1, 8, False),
Im2ColPack(7, 4, 1, 16, False),
]
def test_conv2d_nchw():
cost_all = []
cost_data = []
cost_kernel = []
cost_conv = []
schedules = []
# g1mb1_ic3ih224iw224_oc64oh112ow112_kh7kw7_sh2sw2_ph3pw3_n"resnet18_0" : 0.773982 ms
# in_channel, in_size, num_filter, kernel, stride, padding = 3, 224, 64, 7, 2, 3
# g1mb1_ic64ih56iw56_oc64oh56ow56_kh3kw3_sh1sw1_ph1pw1_n"resnet18_1" : 0.72514 ms
# in_channel, in_size, num_filter, kernel, stride, padding = 64, 56, 64, 3, 1, 1
# g1mb1_ic64ih56iw56_oc64oh56ow56_kh1kw1_sh1sw1_ph0pw0_n"resnet18_2"
# in_channel, in_size, num_filter, kernel, stride, padding = 64, 56, 64, 1, 1, 0
# g1mb1_ic64ih56iw56_oc128oh28ow28_kh3kw3_sh2sw2_ph1pw1_n"resnet18_3"
# in_channel, in_size, num_filter, kernel, stride, padding = 64, 56, 128, 3, 2, 1
# g1mb1_ic64ih56iw56_oc128oh28ow28_kh1kw1_sh2sw2_ph0pw0_n"resnet18_4"
in_channel, in_size, num_filter, kernel, stride, padding = 64, 56, 128, 1, 2, 0
# g1mb1_ic128ih28iw28_oc128oh28ow28_kh3kw3_sh1sw1_ph1pw1_n"resnet18_5"
# in_channel, in_size, num_filter, kernel, stride, padding = 128, 28, 128, 3, 1, 1
# g1mb1_ic128ih28iw28_oc256oh14ow14_kh3kw3_sh2sw2_ph1pw1_n"resnet18_6"
# in_channel, in_size, num_filter, kernel, stride, padding = 128, 28, 256, 3, 2, 1
# g1mb1_ic128ih28iw28_oc256oh14ow14_kh1kw1_sh2sw2_ph0pw0_n"resnet18_7"
# in_channel, in_size, num_filter, kernel, stride, padding = 128, 28, 256, 1, 2, 0
# g1mb1_ic256ih14iw14_oc256oh14ow14_kh3kw3_sh1sw1_ph1pw1_n"resnet18_8"
# in_channel, in_size, num_filter, kernel, stride, padding = 256, 14, 256, 3, 1, 1
# g1mb1_ic256ih14iw14_oc512oh7ow7_kh3kw3_sh2sw2_ph1pw1_n"resnet18_9"
# in_channel, in_size, num_filter, kernel, stride, padding = 256, 14, 512, 3, 2, 1
# g1mb1_ic256ih14iw14_oc512oh7ow7_kh1kw1_sh2sw2_ph0pw0_n"resnet18_10"
# in_channel, in_size, num_filter, kernel, stride, padding = 256, 14, 512, 1, 2, 0
# g1mb1_ic512ih7iw7_oc512oh7ow7_kh3kw3_sh1sw1_ph1pw1_n"resnet18_11"
# in_channel, in_size, num_filter, kernel, stride, padding = 512, 7, 512, 3, 1, 1
out_size = (in_size + 2 * padding - kernel) // stride + 1
vhw_candidates = factors(out_size)
vc_candidates = factors(num_filter)
ba_candidates = factors(in_size)
bc_candidates = factors(kernel)
with open('report/conv_search.txt', 'a') as report:
while True:
vhw_idx = np.random.randint(0, len(vhw_candidates), size=(2,))
vc_idx = np.random.randint(0, len(vc_candidates))
ba_idx = np.random.randint(0, len(ba_candidates))
bc_idx = np.random.randint(0, len(bc_candidates))
sch = SpatialPack(vh=vhw_candidates[vhw_idx[0]], vw=vhw_candidates[vhw_idx[1]],
vc=vc_candidates[vc_idx], ba=ba_candidates[ba_idx], bc=bc_candidates[bc_idx],
unroll=bool(np.random.randint(0, 2)))
try:
print("Trying " + str(sch))
cost = verify_conv2d_nchw_all(sch, 1, in_channel, in_size, num_filter, kernel, stride, padding)
data_cost, kernel_cost, conv_cost = verify_conv2d_nchw(sch, 1, in_channel, in_size, num_filter, kernel, stride, padding)
print("Successful try with %s, conv total time = %f" % (str(sch), cost))
cost_all.append(cost)
cost_data.append(data_cost)
cost_kernel.append(kernel_cost)
cost_conv.append(conv_cost)
schedules.append(sch)
if len(cost_conv) >= 20:
idx = np.argmin(cost_conv)
report.write('%s\tall=%f\tconv=%f\tdata=%f\tkernel=%f\n' %
(str(sch), cost_all[idx], cost_conv[idx], cost_data[idx], cost_kernel[idx]))
report.flush()
cost_all = []
cost_conv = []
cost_kernel = []
cost_data = []
except Exception as e:
print(e)
from __future__ import absolute_import as _abs
import tvm
from tvm import target as _target
from topi import tag
from topi.nn.conv2d import conv2d, _get_schedule
from topi.nn.conv2d import SpatialPack, Im2ColPack
from topi.nn.conv2d import _WORKLOADS, _SCH_TO_DECL_FUNC
from topi.nn.conv2d import _get_workload
from topi.nn.util import infer_pad, infer_stride
from topi import generic
_SCHEDULES = [
# float32 imagenet
SpatialPack(vh=28, vw=8, vc=16, ba=224, bc=1, unroll=True),
SpatialPack(vh=7, vw=1, vc=16, ba=64, bc=4, unroll=True),
SpatialPack(vh=4, vw=7, vc=16, ba=14, bc=1, unroll=True),
SpatialPack(vh=2, vw=7, vc=16, ba=56, bc=1, unroll=True),
SpatialPack(vh=1, vw=14, vc=32, ba=2, bc=1, unroll=False),
SpatialPack(vh=1, vw=7, vc=16, ba=2, bc=1, unroll=False),
SpatialPack(vh=7, vw=2, vc=16, ba=7, bc=1, unroll=False),
SpatialPack(vh=1, vw=7, vc=32, ba=28, bc=1, unroll=False),
SpatialPack(vh=1, vw=7, vc=16, ba=14, bc=1, unroll=True),
SpatialPack(vh=1, vw=7, vc=32, ba=1, bc=1, unroll=False),
SpatialPack(vh=1, vw=7, vc=32, ba=7, bc=1, unroll=False),
SpatialPack(vh=1, vw=7, vc=8, ba=1, bc=1, unroll=True),
# float32 mobilenet
# TODO: need to search for best performance
SpatialPack(2, 2, 4, 28, 1, True),
SpatialPack(1, 4, 8, 14, 1, False),
SpatialPack(1, 2, 16, 8, 1, True),