yzh119

Retrospective on SparseTIR artifact

Once SparseTIR was accepted to ASPLOS 2023, we began constructing the sparsetir-artifact repository for artifact evaluation. Though we already have lots of benchmarking scripts, we found it still not a trivial job to put them together and evaluate in a unified manner. While preparing our artifact, we also found some problems with our profiler and bugs in existing implementations. We carefully addressed these issues and standardized the settings for all baselines. We are writing this post to document the challenges we faced and the lessons we learned from creating the artifact. We aim to provide insight that will benefit researchers and engineers working in related fields.

Notes on performance difference

If you previously read our manuscript on ArXiv, you may have noticed that there are some discrepancies in the reported performance between SparseTIRv3 and our [camera-ready ver

yzh119

import os
import sys
import glob
import logging

logging.basicConfig(level=logging.INFO)


def fix_cite_format(line_number: int, line: str):
    out = ""

yzh119

import tvm
from tvm import tir
from tvm.script import ty
from tvm.tir.schedule.schedule import Schedule


@tvm.script.tir
def ell_spmm(indices_: ty.handle, a_data: ty.handle, b: ty.handle, c: ty.handle) -> None:
    mb = tir.var('int32')
    n = tir.var('int32')

yzh119

import tvm
from tvm import tir
from tvm.script import ty

@tvm.script.tir
def csr_spmm(indptr_: ty.handle, indices_: ty.handle, a_data: ty.handle, b: ty.handle, c: ty.handle) -> None:
    m = tir.var('int32')
    n = tir.var('int32')
    k = tir.var('int32')
    nnz = tir.var('int32')

yzh119

prog = "prog = {:c}{}{:c}{:c}print(prog.format(34, prog, 34, 10))"
print(prog.format(34, prog, 34, 10))

yzh119

#include <stdio.h>
char str[] = "#include <stdio.h>%cchar str[] = %c%s%c;%cint main() {%c    printf(str, 10, 34, str, 34, 10, 10, 10);%c}";
int main() {
	printf(str, 10, 34, str, 34, 10, 10, 10);
}

yzh119

import dgl
import dgl.ops as ops
import numpy as np
import torch as th
import torch.nn as nn

class FFN(nn.Module):
    def __init__(self, d_feat, d_ffn, dropout=0.1):
        super().__init__()
        self.linear_0 = nn.Linear(d_feat, d_ffn)

yzh119

"""Training graphsage w/ fp16.

Usage:

python train_full.py --gpu 0 --fp16 --dataset

Note that GradScaler is not acitvated because the model successfully converges
without gradient scaling.

DGL's Message Passing APIs are not compatible with fp16 yet, hence we disabled

yzh119

"""
This code was modified from the GCN implementation in DGL examples.

Simplifying Graph Convolutional Networks
Paper: https://arxiv.org/abs/1902.07153
Code: https://github.com/Tiiiger/SGC

SGC implementation in DGL.
"""
import argparse, time, math

yzh119

# This part for jupyter notebook setting (if you wants to save, don't use this)
# %matplotlib inline
# %config InlineBackend.figure_format = 'svg'
# import numpy as np
# import matplotlib.pyplot as plt
# plt.rcParams["animation.html"] = "jshtml"


import networkx as nx
from networkx.algorithms import bipartite