Auto-scheduling a Neural Network for mali GPU

Author: Zhao Wu

Auto-tuning for specific devices and workloads is critical for getting the best performance. This is a tutorial on how to tune a whole neural network for mali GPU with the auto-scheduler.

To auto-tune a neural network, we partition the network into small subgraphs and tune them independently. Each subgraph is treated as one search task. A task scheduler slices the time and dynamically allocates time resources to these tasks. The task scheduler predicts the impact of each task on the end-to-end execution time and prioritizes the one that can reduce the execution time the most.

For each subgraph, we use the compute declaration in tvm/python/topi to get the computational DAG in the tensor expression form. We then use the auto-scheduler to construct a search space of this DAG and search for good schedules (low-level optimizations).

Different from the template-based autotvm which relies on manual templates to define the search space, the auto-scheduler does not require any schedule templates. In other words, the auto-scheduler only uses the compute declarations in tvm/python/topi and does not use existing schedule templates.

Note that this tutorial will not run on Windows or recent versions of macOS. To get it to run, you will need to wrap the body of this tutorial in a if __name__ == "__main__": block.

import os
import sys

import numpy as np

import tvm
from tvm import relay, auto_scheduler
import tvm.relay.testing
from tvm.contrib import graph_executor

Define a Network

First, we need to define the network with relay frontend API. We can load some pre-defined network from tvm.relay.testing. We can also load models from MXNet, ONNX, PyTorch, and TensorFlow (see front end tutorials).

For convolutional neural networks, although auto-scheduler can work correctly with any layout, we found the best performance is typically achieved with NHWC layout. We also implemented more optimizations for NHWC layout with the auto-scheduler. So it is recommended to convert your models to NHWC layout to use the auto-scheduler. You can use ConvertLayout pass to do the layout conversion in TVM.

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)

    return mod, params, input_shape, output_shape


# Define the neural network and compilation target.
network = "mobilenet"
batch_size = 1
layout = "NHWC"
# Set this to True if you use ndk tools for cross compiling
use_ndk = True
# Path to cross compiler
os.environ["TVM_NDK_CC"] = "/usr/bin/aarch64-linux-gnu-g++"
target = tvm.target.Target("opencl -device=mali", host="llvm -mtriple=aarch64-linux-gnu")
dtype = "float32"
log_file = "%s-%s-B%d-%s.json" % (network, layout, batch_size, target.kind.name)

Start an RPC Tracker and Register Devices to the Tracker

Please refer to the “Start RPC Tracker” and “Register Devices to RPC Tracker” setions in this tutorial to start an RPC tracker and register devices to the tracker.

# Replace this with the device key in your tracker
device_key = "rk3399"

Extract Search Tasks

Next, we extract the search tasks and their weights from a network. The weight of a task is the number of appearances of the task’s subgraph in the whole network. By using the weight, we can approximate the end-to-end latency of the network as sum(latency[t] * weight[t]), where latency[t] is the latency of a task and weight[t] is the weight of the task. The task scheduler will just optimize this objective.

# 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)
Extract tasks...
========== Task 0  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 56, 56, 128], [1, 1, 128, 128], [1, 1, 1, 128], [1, 56, 56, 128]]) ==========
p0 = PLACEHOLDER [1, 56, 56, 128]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 128, 128]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 128]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 1  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 7, 7, 1024], [1, 1, 1024, 1024], [1, 1, 1, 1024], [1, 7, 7, 1024]]) ==========
p0 = PLACEHOLDER [1, 7, 7, 1024]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 1024, 1024]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 1024]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 2  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 14, 14, 512], [1, 1, 512, 512], [1, 1, 1, 512], [1, 14, 14, 512]]) ==========
p0 = PLACEHOLDER [1, 14, 14, 512]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 512, 512]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 512]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 3  (workload key: ["8c53ca2904398da2889aa7508082d7bb", [1, 7, 7, 1024], [1, 1, 1, 1024]]) ==========
p0 = PLACEHOLDER [1, 7, 7, 1024]
adaptive_pool_sum(ax0, ax1, ax2, ax3) += p0[ax0, ((ax1*7) + rv0), ((ax2*7) + rv1), ax3]
adaptive_pool_avg(ax0, ax1, ax2, ax3) = (adaptive_pool_sum[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 4  (workload key: ["98cde4888c94ec7beaa9972f806856d0", [1, 28, 28, 256], [3, 3, 256, 1], [1, 1, 1, 256], [1, 28, 28, 256]]) ==========
p0 = PLACEHOLDER [1, 28, 28, 256]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 256, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 256]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 5  (workload key: ["08f7449d79e570b7274174709e5e5e01", [1, 1024], [1000, 1024], [1, 1000], [1, 1000]]) ==========
p0 = PLACEHOLDER [1, 1024]
p1 = PLACEHOLDER [1000, 1024]
T_matmul_NT(i0, i1) += (p0[i0, k]*p1[i1, k])
p2 = PLACEHOLDER [1, 1000]
T_add(ax0, ax1) = (T_matmul_NT[ax0, ax1] + p2[ax0, ax1])

========== Task 6  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 112, 112, 32], [1, 1, 32, 64], [1, 1, 1, 64], [1, 112, 112, 64]]) ==========
p0 = PLACEHOLDER [1, 112, 112, 32]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 32, 64]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 64]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 7  (workload key: ["98cde4888c94ec7beaa9972f806856d0", [1, 56, 56, 128], [3, 3, 128, 1], [1, 1, 1, 128], [1, 56, 56, 128]]) ==========
p0 = PLACEHOLDER [1, 56, 56, 128]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 128, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 128]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 8  (workload key: ["88a2e34d300a6ccfcf0228f0b90f13ec", [1, 14, 14, 512], [3, 3, 512, 1], [1, 1, 1, 512], [1, 7, 7, 512]]) ==========
p0 = PLACEHOLDER [1, 14, 14, 512]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 512, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 512]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 9  (workload key: ["7d79c516e212fe1d73f5dbb90eaca2cf", [1, 1000], [1, 1000]]) ==========
p0 = PLACEHOLDER [1, 1000]
T_softmax_maxelem(i0) max= p0[i0, k]
T_softmax_exp(i0, i1) = tir.exp((p0[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 10  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 28, 28, 128], [1, 1, 128, 256], [1, 1, 1, 256], [1, 28, 28, 256]]) ==========
p0 = PLACEHOLDER [1, 28, 28, 128]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 128, 256]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 256]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 11  (workload key: ["98cde4888c94ec7beaa9972f806856d0", [1, 112, 112, 32], [3, 3, 32, 1], [1, 1, 1, 32], [1, 112, 112, 32]]) ==========
p0 = PLACEHOLDER [1, 112, 112, 32]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 113)) && (i2 >= 1)) && (i2 < 113)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 32, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 32]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 12  (workload key: ["88a2e34d300a6ccfcf0228f0b90f13ec", [1, 28, 28, 256], [3, 3, 256, 1], [1, 1, 1, 256], [1, 14, 14, 256]]) ==========
p0 = PLACEHOLDER [1, 28, 28, 256]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 29)) && (i2 >= 1)) && (i2 < 29)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 256, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 256]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 13  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 56, 56, 64], [1, 1, 64, 128], [1, 1, 1, 128], [1, 56, 56, 128]]) ==========
p0 = PLACEHOLDER [1, 56, 56, 64]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 64, 128]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 128]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 14  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 7, 7, 512], [1, 1, 512, 1024], [1, 1, 1, 1024], [1, 7, 7, 1024]]) ==========
p0 = PLACEHOLDER [1, 7, 7, 512]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 512, 1024]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 1024]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 15  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 14, 14, 256], [1, 1, 256, 512], [1, 1, 1, 512], [1, 14, 14, 512]]) ==========
p0 = PLACEHOLDER [1, 14, 14, 256]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 256, 512]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 512]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 16  (workload key: ["88a2e34d300a6ccfcf0228f0b90f13ec", [1, 56, 56, 128], [3, 3, 128, 1], [1, 1, 1, 128], [1, 28, 28, 128]]) ==========
p0 = PLACEHOLDER [1, 56, 56, 128]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 57)) && (i2 >= 1)) && (i2 < 57)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 128, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 128]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 17  (workload key: ["88a2e34d300a6ccfcf0228f0b90f13ec", [1, 112, 112, 64], [3, 3, 64, 1], [1, 1, 1, 64], [1, 56, 56, 64]]) ==========
p0 = PLACEHOLDER [1, 112, 112, 64]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 113)) && (i2 >= 1)) && (i2 < 113)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 64, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, ((i*2) + di), ((j*2) + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 64]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 18  (workload key: ["98cde4888c94ec7beaa9972f806856d0", [1, 7, 7, 1024], [3, 3, 1024, 1], [1, 1, 1, 1024], [1, 7, 7, 1024]]) ==========
p0 = PLACEHOLDER [1, 7, 7, 1024]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 8)) && (i2 >= 1)) && (i2 < 8)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 1024, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 1024]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 19  (workload key: ["ad24d4d2f83975ff580a4833fbf3ef94", [1, 224, 224, 3], [3, 3, 3, 32], [1, 112, 1, 1], [1, 112, 1, 1], [1, 112, 112, 32]]) ==========
p0 = PLACEHOLDER [1, 224, 224, 3]
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 225)) && (i2 >= 1)) && (i2 < 225)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 3, 32]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, ((yy*2) + ry), ((xx*2) + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 112, 1, 1]
T_multiply(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3]*p2[ax0, ax1, 0, 0])
p3 = PLACEHOLDER [1, 112, 1, 1]
T_add(ax0, ax1, ax2, ax3) = (T_multiply[ax0, ax1, ax2, ax3] + p3[ax0, ax1, 0, 0])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 20  (workload key: ["98cde4888c94ec7beaa9972f806856d0", [1, 14, 14, 512], [3, 3, 512, 1], [1, 1, 1, 512], [1, 14, 14, 512]]) ==========
p0 = PLACEHOLDER [1, 14, 14, 512]
PaddedInput(i0, i1, i2, i3) = tir.if_then_else(((((i1 >= 1) && (i1 < 15)) && (i2 >= 1)) && (i2 < 15)), p0[i0, (i1 - 1), (i2 - 1), i3], 0f)
p1 = PLACEHOLDER [3, 3, 512, 1]
DepthwiseConv2d(b, i, j, c) += (PaddedInput[b, (i + di), (j + dj), c]*p1[di, dj, c, 0])
p2 = PLACEHOLDER [1, 1, 1, 512]
T_add(ax0, ax1, ax2, ax3) = (DepthwiseConv2d[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

========== Task 21  (workload key: ["6d628209072e3e3dd8f49359935acea6", [1, 28, 28, 256], [1, 1, 256, 256], [1, 1, 1, 256], [1, 28, 28, 256]]) ==========
p0 = PLACEHOLDER [1, 28, 28, 256]
pad_temp(i0, i1, i2, i3) = p0[i0, i1, i2, i3]
p1 = PLACEHOLDER [1, 1, 256, 256]
conv2d_nhwc(nn, yy, xx, ff) += (pad_temp[nn, (yy + ry), (xx + rx), rc]*p1[ry, rx, rc, ff])
p2 = PLACEHOLDER [1, 1, 1, 256]
T_add(ax0, ax1, ax2, ax3) = (conv2d_nhwc[ax0, ax1, ax2, ax3] + p2[ax0, 0, 0, ax3])
T_relu(ax0, ax1, ax2, ax3) = max(T_add[ax0, ax1, ax2, ax3], 0f)

Note

How to get the hardware parameters from remote device

from tvm.auto_scheduler.utils import request_remote
remote = request_remote(device_key, "127.0.0.1", 9190)
dev = remote.cl()
max_shared_memory_per_block = dev.max_shared_memory_per_block
# There is no explicit local memory limition
# so we can use INT32_MAX to disable the check on local_memory.
max_local_memory_per_block = 2147483647 # INT32_MAX
max_threads_per_block = dev.max_threads_per_block
max_vthread_extent = int(dev.warp_size / 4) if int(dev.warp_size / 4) > 1 else dev.warp_size
warp_size = dev.warp_size
hardware_params = auto_scheduler.HardwareParams(-1, 16, 64,
                                                max_shared_memory_per_block, max_local_memory_per_block,
                                                max_threads_per_block, max_vthread_extent, warp_size)

Now you could pass it to search task and tune

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

Tuning and Evaluate

Now, we set some options for tuning, launch the search tasks and evaluate the end-to-end performance

  • num_measure_trials is the number of measurement trials we can use during the tuning. You can set it to a small number (e.g., 200) for a fast demonstrative run. In practice, we recommend setting it around 800 * len(tasks), which is typically enough for the search to converge. For example, there are 29 tasks in resnet-50, so we can set it as 20000. You can adjust this parameter according to your time budget.

  • In addition, we use RecordToFile to dump measurement records into a log file, The measurement records can be used to query the history best, resume the search, and do more analyses later.

  • see auto_scheduler.TuningOptions, auto_scheduler.LocalRunner for more parameters.

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
        builder=auto_scheduler.LocalBuilder(build_func="ndk" if use_ndk else "default"),
        runner=auto_scheduler.RPCRunner(
            device_key, host="127.0.0.1", port=9190, repeat=3, timeout=50
        ),
        measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
    )

    tuner.tune(tune_option)

    # Compile the whole network
    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, params=params)

    # Create graph executor
    print("=============== Request Remote ===============")
    from tvm.auto_scheduler.utils import request_remote

    remote = request_remote(device_key, "127.0.0.1", 9190)
    dev = remote.cl()
    from tvm.contrib import utils, ndk

    temp = utils.tempdir()
    filename = "deploy_lib.so"
    path_lib = temp.relpath(filename)
    lib.export_library(path_lib, fcompile=ndk.create_shared)
    remote.upload(path_lib)
    loaded_lib = remote.load_module(filename)
    module = graph_executor.GraphModule(loaded_lib["default"](dev))
    data = (np.random.uniform(size=input_shape)).astype(dtype)
    data_tvm = tvm.nd.array(data)
    module.set_input("data", data_tvm)

    # Evaluate
    print("Evaluate inference time cost...")
    print(module.benchmark(dev, repeat=3, min_repeat_ms=500))


# We do not run the tuning in our webpage server since server doesn't have mali gpu.
# Uncomment the following line to run it by yourself.

# tune_and_evaluate()

Note

Explain the printed information during tuning

During the tuning, a lot of information will be printed on the console. They are used for debugging purposes. The most important info is the output of the task scheduler. The following table is a sample output.

----------------------------------------------------------------------
------------------------------  [ Task Scheduler ]
----------------------------------------------------------------------
|  ID  | Latency (ms) | Speed (GFLOPS) | Trials |
-------------------------------------------------
|    0 |        0.010 |           0.40 |     64 |
|    1 |        0.087 |          47.19 |     64 |
|    2 |        0.008 |          -0.00 |     64 |
|    3 |        0.177 |         582.07 |     64 |
|    4 |        0.268 |         862.37 |    256 |
|    5 |        0.166 |         621.13 |    128 |
|    6 |        0.170 |         605.10 |    128 |
|    7 |        0.128 |         403.20 |     64 |
|    8 |        0.189 |         545.71 |     64 |
|    9 |        0.231 |        1001.01 |    448 |
|   10 |        0.155 |         664.80 |    256 |
|   11 |        0.155 |         662.86 |    256 |
|   12 |        0.119 |         434.08 |     64 |
|   13 |        0.199 |         522.13 |     64 |
|   14 |        0.235 |         986.56 |    320 |
|   15 |        0.149 |         689.13 |    128 |
|   16 |        0.155 |         664.80 |    192 |
|   17 |        0.151 |         340.64 |     64 |
|   18 |        0.176 |         597.55 |    128 |
|   19 |        0.220 |        1054.37 |    192 |
|   20 |        0.150 |         686.01 |    128 |
|   21 |        0.159 |         650.88 |    128 |
|   22 |        0.073 |         358.19 |     64 |
|   23 |        0.031 |          70.63 |     64 |
|   24 |        0.251 |         947.73 |    128 |
|   25 |        0.157 |         652.47 |    128 |
|   26 |        0.215 |         954.84 |    128 |
|   27 |        0.237 |         868.92 |    128 |
|   28 |        0.266 |         774.06 |    128 |
-------------------------------------------------
Estimated total latency: 10.016 ms      Trials: 3992    Used time : 1131 s      Next ID: 15

This table lists the latency and (estimated) speed of all tasks. It also lists the allocation of measurement trials for all tasks. The last line prints the total weighted latency of these tasks, which can be a rough estimation of the end-to-end execution time of the network. The last line also prints the total number of measurement trials, total time spent on auto-tuning and the id of the next task to tune.

There will also be some “tvm::Error”s errors, because the auto-scheduler will try some invalid schedules. You can safely ignore them if the tuning can continue, because these errors are isolated from the main process.

Note

Terminate the tuning earlier

You can terminate the tuning earlier by forcibly killing this process. As long as you get at least one valid schedule for each task in the log file, you should be able to do the compilation (the secion below).

Other Tips

  1. During the tuning, the auto-scheduler needs to compile many programs and extract feature from them. This part is CPU-intensive, so a high-performance CPU with many cores is recommended for faster search.

  2. You can use python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json to distill the large log file and only save the best useful records.

  3. You can resume a search from the previous log file. You just need to add a new argument load_log_file when creating the task scheduler in function run_tuning. Say, tuner = auto_scheduler.TaskScheduler(tasks, task_weights, load_log_file=log_file)

  4. If you have multiple target GPUs, you can use all of them for measurements to parallelize the measurements. Check this section to learn how to use the RPC Tracker and RPC Server. To use the RPC Tracker in auto-scheduler, replace the runner in TuningOptions with auto_scheduler.RPCRunner.

Gallery generated by Sphinx-Gallery