【TVM教程】为 NVIDIA GPU 主动调度神经网络
Apache TVM 是一个深度的深度学习编译框架，适用于 CPU、GPU 和各种机器学习加快芯片。更多 TVM 中文文档可访问 →https://tvm.hyper.ai/作者：Lianmin Zheng
针对特定装备和工作负载的主动调优对于获得最佳性能至关重要。本文介绍如何使用 auto-scheduler 为 NVIDIA GPU 调优整个神经网络。
为主动调优神经网络，需要将网络划分为小的子图并独立调优。每个子图被视为一个搜刮任务，任务调度器对时间举行切片并动态地为这些任务分配时间资源，并预测每个任务对端到端执行时间的影响，优先思量最能淘汰执行时间的任务。
对于每个子图，使用 tvm/python/topi 中的盘算声明来获取张量表达式情势的盘算 DAG。然后用 auto-scheduler 来构建这个 DAG 的搜刮空间，并搜刮合适的调度（低级优化）。
与基于 template 的 AutoTVM（依赖手动 template 来定义搜刮空间的） 不同，auto-scheduler 无需任何调度 template。换言之，auto-scheduler 只使用 tvm/python/topi 中的盘算声明，不使用现有的调度 template。
注意，本教程无法在 Windows 或最新版本的 macOS 上运行。如需运行，请将本教程的主体放在 if __name__ == "__main__": 代码块中。
import numpy as np
import tvm
from tvm import relay, auto_scheduler
import tvm.relay.testing
from tvm.contrib import graph_executor
定义网络
起首，要用 Relay 前端 API 定义网络。可以从 tvm.relay.testing 加载一些预定义的网络。也可以从 MXNet、ONNX、PyTorch 和 TensorFlow 加载模子（参见 前端教程）。
对于卷积神经网络，尽管 auto-scheduler 可以在任何布局下正常运行，但通过 NHWC 布局实现的性能最佳。auto-scheduler 对 NHWC 布局举行了很多优化，因此推荐将模子转换为 NHWC 布局，从而得以使用 auto-scheduler。可用 ConvertLayout pass 在 TVM 中举行布局转换。
def get_network(name, batch_size, layout="NHWC", dtype="float32"):
"""Get the symbol definition and random weight of a network"""
# auto-scheduler 更适合 NHWC 布局
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("-"))
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("-"))
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":
# MXNet 模型的示例
from mxnet.gluon.model_zoo.vision import get_model
assert layout == "NCHW"
block = get_model("resnet18_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
# 定义神经网络和编译目标
network = "resnet-18"
batch_size = 1
layout = "NHWC"
target = tvm.target.Target("cuda")
dtype = "float32"
log_file = "%s-%s-B%d-%s.json" % (network, layout, batch_size, target.kind.name)
提取搜刮任务
接下来，从网络中提取搜刮任务及其权重。任务的权重是任务的子图在整个网络中出现的次数。通过使用权重，可以将网络的端到端耽误近似为 sum(latency * weight)，其中 latency 是任务的耽误，而 weight 是任务的权重，任务调度器仅针对该目标举行优化。
# 从网络中提取任务
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)
输出结果：
Extract tasks...
/workspace/python/tvm/driver/build_module.py:268: UserWarning: target_host parameter is going to be deprecated. Please pass in tvm.target.Target(target, host=target_host) instead.
"target_host parameter is going to be deprecated. "
========== Task 0 (workload key: ["8654f16aeddf785bad9f028164b3a48d", , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = placeholder
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
========== Task 1 (workload key: ["c4500b4e2fd04e695c32d2f31bbdc14a", , , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (T_add + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 2 (workload key: ["06f578e6519a86e85028eecf4de64b25", , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = placeholder
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
========== Task 3 (workload key: ["b8b52b9be9df6102466a22a014c44c1f", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 4 (workload key: ["e4cdf917b876dbdd64488c3818d9c141", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 5 (workload key: ["d730bcd28f0920f6b97245e2a11bd8d6", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 8)) && (i2 >= 1)) && (i2 < 8)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
========== Task 6 (workload key: ["b818b53148cd450f86569dfc3e04cb8a", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 6) == 5)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 6) == 4)),..(OMITTED).. (floormod(j, 6) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 6) == 0)), 1f, 0f))))))))))))))))))))))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 4) == 2)),..(OMITTED).. 6) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 7 (workload key: ["ad6cecbf5d85cb1cda3c2bb7af170211", , , , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 8)) && (i2 >= 1)) && (i2 < 8)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
placeholder = PLACEHOLDER
T_multiply(ax0, ax1, ax2, ax3) = (T_add*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (T_multiply + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 8 (workload key: ["f3b6c10fcc6ce01ff01add933e4d21e9", , , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (T_add + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 9 (workload key: ["d7b65649a4dd54becea0a52aabbc5af5", , ]) ==========
placeholder = PLACEHOLDER
T_softmax_maxelem(i0) max= placeholder
T_softmax_exp(i0, i1) = tir.exp((placeholder - T_softmax_maxelem))
T_softmax_expsum(i0) += T_softmax_exp
T_softmax_norm(i0, i1) = (T_softmax_exp/T_softmax_expsum)
========== Task 10 (workload key: ["69115f188984ae34ede37c3b8ca40b43", , ]) ==========
placeholder = PLACEHOLDER
tensor(ax0, ax1, ax2, ax3) += placeholder
tensor(ax0, ax1, ax2, ax3) = (tensor/(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 11 (workload key: ["3a69f9fbc63760d99e36b4c17b3bfc57", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 8)) && (i2 >= 1)) && (i2 < 8)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 12 (workload key: ["06f578e6519a86e85028eecf4de64b25", , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = placeholder
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
========== Task 13 (workload key: ["96daaa9daa1b41bc383b7c05ce8b58de", , , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder, 0f)
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 14 (workload key: ["dac19035dd5fe9424ee8617421b9c817", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
========== Task 15 (workload key: ["96daaa9daa1b41bc383b7c05ce8b58de", , , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), placeholder, 0f)
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 16 (workload key: ["1e3c4211ffd2f2db91078ae4d04b779d", , , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 6) == 5)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 6) == 4)),..(OMITTED).. (floormod(j, 6) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 6) == 0)), 1f, 0f))))))))))))))))))))))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 4) == 2)),..(OMITTED).. 6) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (T_add + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 17 (workload key: ["96daaa9daa1b41bc383b7c05ce8b58de", , , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 3) && (i1 < 227)) && (i2 >= 3)) && (i2 < 227)), placeholder, 0f)
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 18 (workload key: ["3ea73fb9b0364374730d09e068821f95", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 6) == 5)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 6) == 4)),..(OMITTED).. (floormod(j, 6) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 6) == 0)), 1f, 0f))))))))))))))))))))))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 6) == 5) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 6) == 5) && (floormod(j, 4) == 2)),..(OMITTED).. 6) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 6) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
========== Task 19 (workload key: ["d374e472bd9d8164892b9e28a0a8cb59", , , , ]) ==========
placeholder = PLACEHOLDER
data_pad(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), placeholder, 0f)
input_tile(eps, nu, p, ci) = data_pad
B(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 4) == 3)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 4) == 2)),..(OMITTED).. ormod(i, 4) == 0) && (floormod(j, 4) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 4) == 0)), 1f, 0f))))))))))))))))
data_pack(eps, nu, p, ci) += ((input_tile*B)*B)
placeholder = PLACEHOLDER
bgemm(eps, nu, p, co) += (data_pack*placeholder)
A(i, j) = select(((floormod(i, 4) == 3) && (floormod(j, 2) == 1)), 1f, select(((floormod(i, 4) == 3) && (floormod(j, 2) == 0)),..(OMITTED).. ct(((floormod(i, 4) == 0) && (floormod(j, 2) == 1)), 0f, select(((floormod(i, 4) == 0) && (floormod(j, 2) == 0)), 1f, 0f))))))))
inverse(vh, vw, p, co) += ((bgemm*A)*A)
conv2d_winograd(n, h, w, co) = inverse
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_winograd + placeholder)
========== Task 20 (workload key: ["64b98c71af70a904fdbb81d7d4188d84", , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(ax0, ax1, ax2, ax3) = tir.if_then_else(((((ax1 >= 1) && (ax1 < 113)) && (ax2 >= 1)) && (ax2 < 113)), placeholder, -3.40282e+38f)
tensor(ax0, ax1, ax2, ax3) max= pad_temp
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (tensor + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
========== Task 21 (workload key: ["06f578e6519a86e85028eecf4de64b25", , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = placeholder
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
========== Task 22 (workload key: ["7d44c6e3c81cd80f61ff2265b2bae89a", , , , ]) ==========
placeholder = PLACEHOLDER
placeholder = PLACEHOLDER
T_matmul_NT(i, j) += (placeholder*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1) = (T_matmul_NT + placeholder)
========== Task 23 (workload key: ["96daaa9daa1b41bc383b7c05ce8b58de", , , , ]) ==========
placeholder = PLACEHOLDER
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), placeholder, 0f)
placeholder = PLACEHOLDER
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp*placeholder)
placeholder = PLACEHOLDER
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc + placeholder)
T_relu(ax0, ax1, ax2, ax3) = max(T_add, 0f)
开始调优
接下来为调优和启动搜刮任务设置一些选项
[*]measure_ctx 启动不同的测试过程以提供隔离。在测试期间保护主进程免受 GPU 瓦解并制止其他 runtime 冲突。
[*]min_repeat_ms 定义每次测试中一次“重复”的最短持续时间，可以预热 GPU 以获得正确测试结果，通常，推荐设置值 >= 300 ms。
[*]num_measure_trials 是调优期间可以使用的测试次数（根据自己的时间预算调解这个参数），若要快速演示，可将其设置为较小的数字（比方 200）。推荐将其设置为 900 * len(tasks) 左右，以便使搜刮收敛。好比 resnet-18 有 24 个任务，所以可以设置为 20000。
[*]此外，使用 RecordToFile 将测试记录转储到日志文件中，测试记录可用于历史最佳查询、规复搜刮以及举行后续分析。
[*]更多参数参见 auto_scheduler.TuningOptions，auto_scheduler.LocalRunner。
def run_tuning():
print("Begin tuning...")
measure_ctx = auto_scheduler.LocalRPCMeasureContext(repeat=1, min_repeat_ms=300, timeout=10)
tuner = auto_scheduler.TaskScheduler(tasks, task_weights)
tune_option = auto_scheduler.TuningOptions(
num_measure_trials=200, # 将此更改为 20000 以达到最佳性能
runner=measure_ctx.runner,
measure_callbacks=,
)
tuner.tune(tune_option)
# 不在网页服务器中运行调优，因为它需要的时间太长。
# 取消注释运行下面行。
# run_tuning()
备注
解释调优过程中打印的信息
在调优过程中，控制台上会打印很多用于调试的信息，最重要的信息是任务调度步调的输出，下表是输出示例。
------------------------------ [ Task Scheduler ]
| ID | Latency (ms) | Speed (GFLOPS) | Trials |
| 0 | 0.005 | 0.88 | 64 |
| 1 | 0.010 | 99.10 | 64 |
| 2 | 0.006 | 0.00 | 64 |
| 3 | 0.145 | 979.78 | 384 |
| 4 | 0.130 | 1097.02 | 384 |
| 5 | 0.143 | 992.69 | 384 |
| 6 | 0.076 | 1526.86 | 192 |
| 7 | 0.115 | 999.44 | 320 |
| 8 | 0.079 | 1449.39 | 320 |
| 9 | 0.122 | 938.73 | 384 |
| 10 | 0.063 | 1832.98 | 192 |
| 11 | 0.072 | 1763.62 | 256 |
| 12 | 0.062 | 2036.40 | 192 |
| 13 | 0.068 | 1874.44 | 192 |
| 14 | 0.049 | 2346.50 | 128 |
| 15 | 0.076 | 1694.31 | 256 |
| 16 | 0.067 | 1933.30 | 448 |
| 17 | 0.076 | 1680.90 | 256 |
| 18 | 0.022 | 98.43 | 64 |
| 19 | 0.076 | 3112.55 | 192 |
| 20 | 0.013 | 2026.44 | 64 |
| 21 | 0.011 | 1136.69 | 64 |
| 22 | 0.013 | 992.47 | 64 |
| 23 | 0.020 | 627.56 | 64 |
Estimated total latency: 1.587 ms Trials: 4992 Used time : 13296 s Next ID: 3
此表列出了所有任务的耽误和（预估）速度，还列出了所有任务的测试分配。末了一行打印了这些任务的总加权耽误，可以大略估计网络的端到端执行时间。末了一行还打印了测试试验的总数、主动调优所花费的总时间以及下一个要调优的任务的 ID。
还有一些「tvm::Error」错误，因为 auto-scheduler 会尝试一些无效的调度。若调优继承运行，则可以忽略这些错误，因为这些错误与主进程隔离。
备注
提前终止调优
可以通过逼迫终止此进程来提前终止调优，只要在日志文件中为每个任务获得至少一个有效的调度，就能够举行编译（下面的部门）。
编译及评估
主动调优后，用找到的最佳调度来编译网络。在主动调优期间，所有测试记录都被转储到日志文件中，可以读取日志文件加载最佳调度。
# 用历史最佳编译
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)
# 创建图执行器
dev = tvm.device(str(target), 0)
module = graph_executor.GraphModule(lib["default"](dev))
data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))
module.set_input("data", data_tvm)
# 评估
print("Evaluate inference time cost...")
print(module.benchmark(dev, repeat=3, min_repeat_ms=500))
输出结果：
Compile...
/workspace/python/tvm/driver/build_module.py:268: UserWarning: target_host parameter is going to be deprecated. Please pass in tvm.target.Target(target, host=target_host) instead.
"target_host parameter is going to be deprecated. "
Evaluate inference time cost...
Execution time summary:
mean (ms) median (ms) max (ms) min (ms) std (ms)
10.0003 9.9944 10.0327 9.9738 0.0244
其他本事
[*]在调优过程中，auto-scheduler 需要编译很多步调，并从中提取特征。这部门会占用大量 CPU 资源，所以推荐使用多核的高性能 CPU，加快搜刮速度。
[*]可以用 python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json 提取大日志文件，并仅生存最有效的记录。
[*]可以从以前的日志文件规复搜刮，只需要在函数 run_tuning 中创建任务调度步调时添加一个新参数 load_log_file。好比，tuner = auto_scheduler.TaskScheduler(tasks, task_weights, load_log_file=log_file)
[*]若有多个 target CPU，则可以将所有这些 CPU 用于并行化测试。查看这 部门 了解如何使用 RPC 跟踪器和 RPC 服务器。要在 auto-scheduler 中使用 RPC 跟踪器，请将 TuningOptions 中的 runner 更换为 auto_scheduler.RPCRunner。
下载 Python 源代码：tune_network_cuda.py
下载 Jupyter Notebook：tune_network_cuda.ipynb
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。
页:
[1]