ARM CPU自动调度神经网络

对特定设备和工作负载进行自动调度,对于获得最佳性能至关重要。通过RPC使用自动调度器为ARM CPU调度整个神经网络。

为了自动调度神经网络,将网络划分为小的子图,进行独立调度。每个子图被视为一个搜索任务。任务调度程序对时间进行分片,为这些任务动态分配时间资源。任务调度程序预测每个任务对端到端执行时间的影响,确定最大程度地减少执行时间的任务的优先级。

对于每个子图,使用compute声明tvm/python/topi,获取张量表达式形式的计算DAG。使用自动调度器来构造此DAG的搜索空间,搜索良好的调度(低级优化)。

与手动模板定义搜索空间的基于模板的autotvm不同,自动调度程序不需要任何调度模板。换句话说,自动调度程序仅在其中使用计算声明tvm/python/topi,而不使用现有的调度模板。

本文无法在Windows或最新版本的macOS上运行。要使其运行,需要将本文的内容包装在一个if __name__ == "__main__":块中。

import numpy as np

import tvm

from tvm import relay, auto_scheduler

import tvm.relay.testing

from tvm.contrib import graph_runtime

from tvm.contrib.utils import tempdir

定义网络

首先,需要使用Relay中继前端API定义网络。从加载一些预定义的网络tvm.relay.testing。还从MXNet,ONNX,PyTorch和TensorFlow加载模型。

对于卷积神经网络,尽管自动调度程序在任何布局下正常工作,使用NHWC布局通常实现最佳性能。使用自动调度程序对NHWC布局实施了更多优化。建议将模型转换为NHWC布局以使用自动调度程序。使用ConvertLayout传递在TVM中进行布局转换。

def get_network(name, batch_size, layout="NHWC", dtype="float32"):

"""Get the symbol definition and random weight of a network"""

# auto-scheduler prefers NHWC layout

if layout == "NHWC":

image_shape = (224, 224, 3)

elif layout == "NCHW":

image_shape = (3, 224, 224)

else:

raise ValueError("Invalid layout: " + layout)

input_shape = (batch_size,) + image_shape

output_shape = (batch_size, 1000)

if name.startswith("resnet-"):

n_layer = int(name.split("-")[1])

mod, params = relay.testing.resnet.get_workload(

num_layers=n_layer,

batch_size=batch_size,

layout=layout,

dtype=dtype,

image_shape=image_shape,

)

elif name.startswith("resnet3d-"):

n_layer = int(name.split("-")[1])

mod, params = relay.testing.resnet.get_workload(

num_layers=n_layer,

batch_size=batch_size,

layout=layout,

dtype=dtype,

image_shape=image_shape,

)

elif name == "mobilenet":

mod, params = relay.testing.mobilenet.get_workload(

batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape

)

elif name == "squeezenet_v1.1":

assert layout == "NCHW", "squeezenet_v1.1 only supports NCHW layout"

mod, params = relay.testing.squeezenet.get_workload(

version="1.1",

batch_size=batch_size,

dtype=dtype,

image_shape=image_shape,

)

elif name == "inception_v3":

input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)

mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)

elif name == "mxnet":

# an example for mxnet model

from mxnet.gluon.model_zoo.vision import get_model

assert layout == "NCHW"

block = get_model("resnet50_v1", pretrained=True)

mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)

net = mod["main"]

net = relay.Function(

net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs

)

mod = tvm.IRModule.from_expr(net)

return mod, params, input_shape, output_shape

启动RPC跟踪器

TVM使用RPC会话与ARM板进行通信。在调度期间,调谐器会将生成的代码发送到电路板上,测量电路板上的代码速度。

为了扩大调度范围,TVM使用RPC Tracker来管理分布式设备。RPC跟踪器是一个集中式控制器节点。将所有设备注册到跟踪器。例如,如果有10部电话,全部注册到跟踪器,并行运行10次测量,从而加快了调谐过程。

要启动RPC跟踪器,在主机上运行此命令。在整个调度过程中都需要使用跟踪器,因此需要为此命令打开一个新终端:

python -m tvm.exec.rpc_tracker --host=0.0.0.0 --port=9190

预期的输出是

INFO:RPCTracker:bind to 0.0.0.0:9190

将设备注册到RPC跟踪器

现在,将设备注册到跟踪器。第一步是为ARM设备构建TVM运行时。

  • 对于Linux:遵循本节在设备上构建TVM运行时,以在设备上构建TVM运行时。然后通过以下方式将设备注册到跟踪器
  • python -m tvm.exec.rpc_server --tracker=[HOST_IP]:9190 --key=rasp4b-64

(用[HOST_IP]主机的IP地址代替)

  • 对于Android:在Android设备上安装TVM RPC APK。确保通过android rpc测试。已经注册了设备。在调度过程中,必须转到开发人员选项,启用“更改时保持屏幕唤醒”并为手机充电以使其稳定。

注册设备后,通过查询rpc_tracker进行确认

python -m tvm.exec.query_rpc_tracker --host=0.0.0.0 --port=9190

例如,如果有2个Huawei mate10 pro,11个具有64位操作系统的Raspberry Pi 4B和2个rk3399,则输出是

Queue Status

----------------------------------

key          total  free  pending

----------------------------------

mate10pro    2      2     0

rk3399       2      2     0

rasp4b-64    11     11    0

----------------------------------

将多个设备注册到跟踪器,以加快调谐过程中的测量速度。

设置调度选项

调度之前,应该应用一些配置。以带有64位操作系统(Ubuntu 20.04)的Raspberry Pi 4b 4GB主板为例。在设置中,应该相应地修改目标和device_key。如果使用的是Android手机,use_ndk设置为True。

#### DEVICE CONFIG ####

# Replace "aarch64-linux-gnu" with the correct target of your board.

# This target is used for cross compilation. You can query it by :code:`gcc -v` on your device.

# FIXME(tmoreau89, merrymercy): We leave '-device=arm_cpu' out of the target string

#                               because we're sharing x86 op strategy.

target = tvm.target.Target("llvm -mtriple=aarch64-linux-gnu -mattr=+neon")

# Also replace this with the device key in your tracker

device_key = "rasp4b-64"

# Set this to True if you use ndk tools for cross compiling

# And also set the environment variable below to point to the cross compiler

use_ndk = False

# os.environ["TVM_NDK_CC"] = "/usr/bin/aarch64-linux-gnu-g++"

#### TUNING OPTION ####

network = "mobilenet"

batch_size = 1

layout = "NHWC"

dtype = "float32"

log_file = "%s-%s-B%d-%s.json" % (network, layout, batch_size, target.kind.name)

提取搜索任务

接下来,从网络中提取搜索任务及其权重。任务的权重是该任务的子图在整个网络中的出现次数。通过使用权重,将网络的端到端延迟近似为,sum(latency[t] * weight[t])其中是任务的延迟,latency[t]weight[t]是任务的权重。任务调度程序将仅优化此目标。

# Extract tasks from the network

print("Extract tasks...")

mod, params, input_shape, output_shape = get_network(network, batch_size, layout, dtype=dtype)

tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)

for idx, task in enumerate(tasks):

print("========== Task %d  (workload key: %s) ==========" % (idx, task.workload_key))

print(task.compute_dag)

出去:

Extract tasks...

========== Task 0  (workload key: ["d7b65649a4dd54becea0a52aabbc5af5", 1, 1000, 1, 1000]) ==========

placeholder = PLACEHOLDER [1, 1000]

T_softmax_maxelem(i0) max= placeholder[i0, k]

T_softmax_exp(i0, i1) = tir.exp((placeholder[i0, i1] - T_softmax_maxelem[i0]))

T_softmax_expsum(i0) += T_softmax_exp[i0, k]

T_softmax_norm(i0, i1) = (T_softmax_exp[i0, i1]/T_softmax_expsum[i0])

========== Task 1  (workload key: ["9847f8cc0b305137f49f2c5c0c8ab25d", 1, 1024, 1000, 1024, 1000, 1, 1000]) ==========

placeholder = PLACEHOLDER [1, 1024]

placeholder = PLACEHOLDER [1000, 1024]

T_dense(i, j) += (placeholder[i, k]*placeholder[j, k])

placeholder = PLACEHOLDER [1000]

T_add(ax0, ax1) = (T_dense[ax0, ax1] + placeholder[ax1])

========== Task 2  (workload key: ["69115f188984ae34ede37c3b8ca40b43", 1, 7, 7, 1024, 1, 1, 1, 1024]) ==========

placeholder = PLACEHOLDER [1, 7, 7, 1024]

tensor(ax0, ax1, ax2, ax3) += placeholder[ax0, ((ax1*7) + rv0), ((ax2*7) + rv1), ax3]

tensor(ax0, ax1, ax2, ax3) = (tensor[ax0, ax1, ax2, ax3]/(float32((select((bool)1, ((ax1 + 1)*7), (((ax1 + 1)*7) + 1)) - (ax1*7)))*float32((select((bool)1, ((ax2 + 1)*7), (((ax2 + 1)*7) + 1)) - (ax2*7)))))

========== Task 3  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 7, 7, 1024, 1, 1, 1024, 1024, 1, 1, 1, 1024, 1, 7, 7, 1024]) ==========

placeholder = PLACEHOLDER [1, 7, 7, 1024]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 1024, 1024]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 1024]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 4  (workload key: ["06fce76bd84cb904eee50b905ca9449a", 1, 7, 7, 1024, 3, 3, 1024, 1, 1, 1, 1, 1024, 1, 7, 7, 1024]) ==========

placeholder = PLACEHOLDER [1, 7, 7, 1024]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 8)) && (i2 >= 1)) && (i2 < 8)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 1024, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 1024]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 5  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 7, 7, 512, 1, 1, 512, 1024, 1, 1, 1, 1024, 1, 7, 7, 1024]) ==========

placeholder = PLACEHOLDER [1, 7, 7, 512]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 512, 1024]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 1024]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 6  (workload key: ["c87ba68bc180312f5716af09a77ca15b", 1, 14, 14, 512, 3, 3, 512, 1, 1, 1, 1, 512, 1, 7, 7, 512]) ==========

placeholder = PLACEHOLDER [1, 14, 14, 512]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 512, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 512]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 7  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 14, 14, 512, 1, 1, 512, 512, 1, 1, 1, 512, 1, 14, 14, 512]) ==========

placeholder = PLACEHOLDER [1, 14, 14, 512]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 512, 512]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 512]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 8  (workload key: ["06fce76bd84cb904eee50b905ca9449a", 1, 14, 14, 512, 3, 3, 512, 1, 1, 1, 1, 512, 1, 14, 14, 512]) ==========

placeholder = PLACEHOLDER [1, 14, 14, 512]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 512, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 512]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 9  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 14, 14, 256, 1, 1, 256, 512, 1, 1, 1, 512, 1, 14, 14, 512]) ==========

placeholder = PLACEHOLDER [1, 14, 14, 256]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 256, 512]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 512]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 10  (workload key: ["c87ba68bc180312f5716af09a77ca15b", 1, 28, 28, 256, 3, 3, 256, 1, 1, 1, 1, 256, 1, 14, 14, 256]) ==========

placeholder = PLACEHOLDER [1, 28, 28, 256]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 256, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 256]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 11  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 28, 28, 256, 1, 1, 256, 256, 1, 1, 1, 256, 1, 28, 28, 256]) ==========

placeholder = PLACEHOLDER [1, 28, 28, 256]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 256, 256]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 256]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 12  (workload key: ["06fce76bd84cb904eee50b905ca9449a", 1, 28, 28, 256, 3, 3, 256, 1, 1, 1, 1, 256, 1, 28, 28, 256]) ==========

placeholder = PLACEHOLDER [1, 28, 28, 256]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 256, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 256]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 13  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 28, 28, 128, 1, 1, 128, 256, 1, 1, 1, 256, 1, 28, 28, 256]) ==========

placeholder = PLACEHOLDER [1, 28, 28, 128]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 128, 256]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 256]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 14  (workload key: ["c87ba68bc180312f5716af09a77ca15b", 1, 56, 56, 128, 3, 3, 128, 1, 1, 1, 1, 128, 1, 28, 28, 128]) ==========

placeholder = PLACEHOLDER [1, 56, 56, 128]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 128, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 128]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 15  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 56, 56, 128, 1, 1, 128, 128, 1, 1, 1, 128, 1, 56, 56, 128]) ==========

placeholder = PLACEHOLDER [1, 56, 56, 128]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 128, 128]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 128]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 16  (workload key: ["06fce76bd84cb904eee50b905ca9449a", 1, 56, 56, 128, 3, 3, 128, 1, 1, 1, 1, 128, 1, 56, 56, 128]) ==========

placeholder = PLACEHOLDER [1, 56, 56, 128]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 128, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 128]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 17  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 56, 56, 64, 1, 1, 64, 128, 1, 1, 1, 128, 1, 56, 56, 128]) ==========

placeholder = PLACEHOLDER [1, 56, 56, 64]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 64, 128]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 128]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 18  (workload key: ["c87ba68bc180312f5716af09a77ca15b", 1, 112, 112, 64, 3, 3, 64, 1, 1, 1, 1, 64, 1, 56, 56, 64]) ==========

placeholder = PLACEHOLDER [1, 112, 112, 64]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 113)) && (i2 >= 1)) && (i2 < 113)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 64, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 64]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 19  (workload key: ["6b7583cf23c7c37d3212cad9d06e58c1", 1, 112, 112, 32, 1, 1, 32, 64, 1, 1, 1, 64, 1, 112, 112, 64]) ==========

placeholder = PLACEHOLDER [1, 112, 112, 32]

PaddedInput(i0, i1, i2, i3) = placeholder[i0, i1, i2, i3]

placeholder = PLACEHOLDER [1, 1, 32, 64]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, (yy + ry), (xx + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 1, 1, 64]

T_add(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 20  (workload key: ["06fce76bd84cb904eee50b905ca9449a", 1, 112, 112, 32, 3, 3, 32, 1, 1, 1, 1, 32, 1, 112, 112, 32]) ==========

placeholder = PLACEHOLDER [1, 112, 112, 32]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 113)) && (i2 >= 1)) && (i2 < 113)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 32, 1]

DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*placeholder[di, dj, c, 0])

placeholder = PLACEHOLDER [1, 1, 1, 32]

T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + placeholder[ax0, 0, 0, ax3])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 21  (workload key: ["98418eda02701ddd175ad50e364a0638", 1, 224, 224, 3, 3, 3, 3, 32, 1, 112, 1, 1, 1, 112, 1, 1, 1, 112, 112, 32]) ==========

placeholder = PLACEHOLDER [1, 224, 224, 3]

PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 225)) && (i2 >= 1)) && (i2 < 225)), placeholder[i0, (i1 - 1), (i2 - 1), i3], 0f)

placeholder = PLACEHOLDER [3, 3, 3, 32]

Conv2dOutput(nn, yy, xx, ff) += (PaddedInput[nn, ((yy*2) + ry), ((xx*2) + rx), rc]*placeholder[ry, rx, rc, ff])

placeholder = PLACEHOLDER [1, 112, 1, 1]

T_multiply(ax0, ax1, ax2, ax3) = (Conv2dOutput[ax0, ax1, ax2, ax3]*placeholder[ax0, ax1, 0, 0])

placeholder = PLACEHOLDER [1, 112, 1, 1]

T_add(ax0, ax1, ax2, ax3) = (T_multiply[ax0, ax1, ax2, ax3] + placeholder[ax0, ax1, 0, 0])

T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

调优与评估

设置一些选项来优化和启动搜索任务

  • num_measure_trials是在调度期间使用的测量试验的次数。将其设置为较小的数字(例如200)以进行快速演示。实际上,建议将其设置为800 * len(tasks),通常足以使搜索收敛。例如,resnet-50中有29个任务,因此将其设置为20000。根据时间预算调度。
  • 此外,还用RecordToFile将测量记录转储到日志文件中,些测量记录可用于最好地查询历史记录,恢复搜索以及以后进行更多分析。
  • 有关更多参数auto_scheduler.TuningOptions, 参见auto_scheduler.LocalRunner

自动调度后,使用发现的最佳调度来编译网络。在自动调度期间,所有测量记录都将转储到日志文件中,因此读取日志文件并加载最佳调度。

def tune_and_evaluate():

print("Begin tuning...")

tuner = auto_scheduler.TaskScheduler(tasks, task_weights)

tune_option = auto_scheduler.TuningOptions(

num_measure_trials=200,  # change this to 20000 to achieve the best performance

runner=auto_scheduler.RPCRunner(

device_key,

host="0.0.0.0",

port=9191,

timeout=30,

repeat=1,

min_repeat_ms=200,

enable_cpu_cache_flush=True,

),

measure_callbacks=[auto_scheduler.RecordToFile(log_file)],

)

tuner.tune(tune_option)

# Compile with the history best

print("Compile...")

with auto_scheduler.ApplyHistoryBest(log_file):

with tvm.transform.PassContext(

opt_level=3, config={"relay.backend.use_auto_scheduler": True}

):

lib = relay.build(mod, target=target, params=params)

# Export library

tmp = tempdir()

if use_ndk:

from tvm.contrib import ndk

filename = "net.so"

lib.export_library(tmp.relpath(filename), ndk.create_shared)

else:

filename = "net.tar"

lib.export_library(tmp.relpath(filename))

# Upload module to device

print("Upload...")

remote = auto_scheduler.utils.request_remote(device_key, "0.0.0.0", 9191, timeout=10000)

remote.upload(tmp.relpath(filename))

rlib = remote.load_module(filename)

# Create graph runtime

ctx = remote.cpu()

module = graph_runtime.GraphModule(rlib["default"](ctx))

data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))

module.set_input("data", data_tvm)

# Evaluate

print("Evaluate inference time cost...")

ftimer = module.module.time_evaluator("run", ctx, repeat=3, min_repeat_ms=500)

prof_res = np.array(ftimer().results) * 1e3  # convert to millisecond

print(

"Mean inference time (std dev): %.2f ms (%.2f ms)" % (np.mean(prof_res), np.std(prof_res))

)

# We do not run the tuning in our webpage server since the server doesn't have a Raspberry Pi,

# or device tracker running.

# Uncomment the following line to run it by yourself.

# tune_and_evaluate()

笔记

调度期间解释打印的信息

在调度期间,控制台上会打印很多信息。用于调试目的。最重要的信息是任务调度程序的输出。下表是示例输出。

----------------------------------------------------------------------

------------------------------  [ Task Scheduler ]

----------------------------------------------------------------------

|  ID  | Latency (ms) | Speed (GFLOPS) | Trials |

-------------------------------------------------

|    0 |        0.013 |           0.31 |     64 |

|    1 |        0.845 |           2.43 |    448 |

|    2 |        0.046 |          -0.00 |     64 |

|    3 |        4.194 |          24.53 |   2112 |

|    4 |        0.109 |           9.21 |     64 |

|    5 |        1.759 |          29.27 |    896 |

|    6 |        0.083 |           6.01 |     64 |

|    7 |        3.084 |          33.38 |   7680 |

|    8 |        0.136 |          14.78 |    384 |

|    9 |        1.349 |          38.23 |    768 |

|   10 |        0.133 |           7.55 |    128 |

|   11 |        2.747 |          37.56 |   1536 |

|   12 |        0.338 |          11.87 |    192 |

|   13 |        1.295 |          40.00 |    704 |

|   14 |        0.482 |           4.16 |    256 |

|   15 |        2.686 |          38.56 |   1344 |

|   16 |        0.884 |           9.08 |    448 |

|   17 |        1.332 |          39.18 |    704 |

|   18 |        1.045 |           3.84 |    576 |

|   19 |        1.391 |          38.09 |    704 |

|   20 |        0.777 |          10.34 |    448 |

|   21 |        0.739 |          30.97 |    448 |

-------------------------------------------------

Estimated total latency: 38.347 ms      Trials: 19992   Used time : 19260 s     Next ID: 3

下表列出了所有任务的延迟和(估计)速度。列出了所有任务的测量试验分配。最后一行显示这些任务的总加权延迟,粗略估计网络的端到端执行时间。最后一行还显示测量试验的总数,自动调度所花费的总时间,以及要调度的下一个任务的ID。

将出现一些“ dmlc :: Error”错误,因为自动调度程序将尝试一些无效的调度。如果继续进行调度,则放心地忽略,因为这些错误与主要过程是隔离的。

笔记

提前终止调度

通过强制终止此过程来提前终止调度。只要在日志文件中为每个任务至少获得一个有效的调度,就应该能够进行编译(下面的部分)。

其他技巧

  1. 在调度过程中,自动调度器需要编译许多程序并从中提取功能。该部分占用大量CPU,建议使用具有多个内核的高性能CPU,以加快搜索速度。
  2. 用来提取大型日志文件,而仅保存最有用的记录。python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json
  3. 从上一个日志文件继续搜索。load_log_file在function中创建任务调度程序时,只需添加一个新参数run_tuning。

tuner = auto_scheduler.TaskScheduler(tasks, task_weights, load_log_file=log_file)

  1. 如果有多个目标CPU,则全部用于测量以并行化测量。检查本节 以了解如何使用RPC跟踪器和RPC服务器。要在自动调度使用RPC跟踪, 用auto_scheduler.RPCRunner更换转轮TuningOptions。

ARM CPU自动调度神经网络的更多相关文章

  1. 为x86 CPU自动调度神经网络

    为x86 CPU自动调度神经网络 对特定设备和工作负载进行自动调试对于获得最佳性能至关重要.这是有关如何使用自动调度器为x86 CPU调试整个神经网络的文档. 为了自动调试神经网络,将网络划分为小的子 ...

  2. NVIDIA GPU自动调度神经网络

    NVIDIA GPU自动调度神经网络 对特定设备和工作负载进行自动调整对于获得最佳性能至关重要.这是有关如何使用自动调度器为NVIDIA GPU调整整个神经网络. 为了自动调整神经网络,将网络划分为小 ...

  3. NVIDIA GPU的神经网络自动调度

    NVIDIA GPU的神经网络自动调度 针对特定设备和工作负载的自动调整对于获得最佳性能至关重要.这是一个关于如何使用自动调度器为NVIDIA GPU调整整个神经网络的资料. 为了自动调整一个神经网络 ...

  4. CPU的自动调度矩阵乘法

    CPU的自动调度矩阵乘法 这是一个有关如何对CPU使用自动调度程序的文档. 与依靠手动模板定义搜索空间的基于模板的autotvm不同,自动调度程序不需要任何模板.用户只需要编写计算声明,而无需任何调度 ...

  5. TVM自动调度器

    TVM自动调度器 随着模型大小,算子多样性和硬件异构性的不断增长,优化深度神经网络的执行速度非常困难.从计算的角度来看,深度神经网络只是张量计算的一层又一层.这些张量计算(例如matmul和conv2 ...

  6. arm cpu的架构及分类说明

    今天在编译mplayer for mx27ads的时候, 碰到了armv5te与armv6优化的问题. 默认的交叉编译器支持armv5te也支持armv6,就默认使用了mplayer中mpeg4的ar ...

  7. ARM CPU大小端

    ARM CPU大小端: 大端模式:低位字节存在高地址上,高位字节存在低地址上 小端模式:高位字节存在高地址上,低位字节存在低地址上 STM32属于小端模式,简单的说,比如u32 temp=0X1234 ...

  8. ARM CPU与Intel x86 CPU性能比较

    Qualcomm ARM CPU与Intel x86 CPU性能比较 随着移动互联网时代的到来,Qualcomm(高通).Texas Instruments(德州仪器)等基于ARM架构的CPU受到越来 ...

  9. 10 -- 深入使用Spring -- 5... 实现任务的自动调度

    10.5 实现任务的自动调度 10.5.1 使用Quartz 10.5.2 在Spring中使用Quartz

随机推荐

  1. ListBox控件简单的数据绑定

    <ListBox Margin="17,8,15,26" Name="ListBoxName" ItemsSource="{Binding Ta ...

  2. [CTF]ROT5/13/18/47位移密码

    [CTF]ROT5/13/18/47位移密码 ---------------转换网站 https://www.qqxiuzi.cn/bianma/ROT5-13-18-47.php ROT5:只对数字 ...

  3. [CTF]凯撒密码

    [CTF]凯撒密码 ---------------------  作者:___Blue_H  来源:CSDN  原文:https://blog.csdn.net/qq_37653144/article ...

  4. XCTF-supersqli

    supersqli 进来有个输入框,看内容应该是var_dump了sql查询结果 单引号有报错,万能语句能用,注释符#没被ban 打了个union select,给提示ban了一堆关键字,而且忽略大小 ...

  5. 【实用小技巧】spring springmvc集成shiro时报 No bean named 'shiroFilter' available

    查了网上的,很多情况,不同的解决办法,总归一点就是配置文件加载的问题. 先看下配置文件中的配置 web.xml中的主要配置(这是修改后不在报错的:仅仅修改了一个位置:[classpath:spring ...

  6. 第一章 FreeBSD之系统安装

    一.默认选择 >>[Boot Multi User],按回车键 二.默认选择 >> [Install] 三.选择默认的键盘设置 >>[Select] 四.配置主机名 ...

  7. c++debug&注意事项 自用 持续更新

    cin后回车程序直接退出: 加system("pause");在return 0;前面 C++ 控制cout输出的小数位数 C++中的cout.setf().cout.precis ...

  8. Educational Codeforces Round 92 (Rated for Div. 2)

    A.LCM Problem 题意:最小公倍数LCM(x,y),处于[l,r]之间,并且x,y也处于[l,r]之间,给出l,r找出x,y; 思路:里面最小的最小公倍数就是基于l左端点的,而那个最小公倍数 ...

  9. 【BUAA_2020_软工】个人作业

    个人项目作业博客 1. 在文章开头给出教学班级和可克隆的 Github 项目地址(例子如下).(1') 项目 内容 北航2020软工 班级博客 作业要求 具体要求 项目GitHub地址 个人项目 教学 ...

  10. Java_集合之三

    1.红黑树 1.1概述[理解] 红黑树的特点 平衡二叉B树 每一个节点可以是红或者黑 红黑树不是高度平衡的,它的平衡是通过"自己的红黑规则"进行实现的 红黑树的红黑规则有哪些 每一 ...