yaoyaoding

import os
import numpy as np
import tvm
from tvm import autotvm
from tvm import relay
import tvm.relay.testing
from tvm.autotvm.tuner import XGBTuner, GATuner, RandomTuner, GridSearchTuner
from tvm.contrib.util import tempdir
import tvm.contrib.graph_runtime as runtime

yaoyaoding

import numpy as np
import argparse
import os
from typing import Tuple
import io
import tvm
import tvm.relay
import time
import cv2
import torch

yaoyaoding

Names:	GFlops:	MParams:
resnet50	4.113562624	25.557032
resnext50	4.261431296	25.028904
isetnet_0	2.473464832	15.339816
isetnet_1	2.778018304	17.054104
isetnet_2	6.111159296	34.855528
isetnet_3	7.098017792	39.636352
isetnet_4	2.746732544	16.71396
isetnet_5	5.106156032	29.349224
isetnet_6	3.028927488	18.108968

yaoyaoding

import taso

def get_pads(kernel, padding):
    if sum(padding) == 0 and sum(kernel) > 2:
        pads = "VALID"
    else:
        pads = "SAME"
    return pads

def conv2d(graph, v, out_channels, kernel=(1, 1), stride=(1, 1), padding=(0, 0)):

yaoyaoding

import tvm
from tvm import te, tir


def extract_launch_config(prim_func: tir.PrimFunc):
    """
    Extract the launch configuration of given prim_func.

    Parameters
    ----------

yaoyaoding

#include <cassert>
#include <cstdio>
extern "C" {

__device__ __forceinline__ void matmul_bt128x128_bsz256_s128x128_block_c_init_warp(float out[64]) {
  int32_t lane_id = (threadIdx.x % 32);
  out[0] = 0.0;
  out[1] = 0.0;
  out[2] = 0.0;
  out[3] = 0.0;

yaoyaoding

#include <cassert>
#include <cstdio>
extern "C" {

__device__ __forceinline__ void matmul_bt128x128_bsz256_s128x128_block_c_init_warp(float out[64]) {
  int32_t lane_id = (threadIdx.x % 32);
  out[0] = 0.0;
  out[1] = 0.0;
  out[2] = 0.0;
  out[3] = 0.0;

yaoyaoding

from typing import List, Union
import json
from collections import defaultdict


class Model:
    def __init__(self, graph, description="", author="", company="", license="", domain="", source=""):
        self.graphs: List[Graph] = [graph]
        self.description: str = description
        self.author: str = author

yaoyaoding

(Beta) Hidet: a dynamo backend focuses on inference acceleration

With torch dynamo, we can dispatch a pytorch model to other awesome deep learning framework/compilers for acceleration. Hidet is one of such deep learning compilers that accelerates your model with a bunch of optimizations (e.g., subgraph fusion, rewriting and kernel tuning). To use hidet, please first install it via

$ pip install hidet

Then you can enable it via torch.compile(model, backend='hidet') as shown in the code snippet below:

import torch
import hidet 

yaoyaoding

import atexit

# clean the cache dir in envorioment variable TRITON_CACHE_DIR
import os
import shutil

if 'TRITON_CACHE_DIR' not in os.environ:
    os.environ['TRITON_CACHE_DIR'] = './triton_cache'
cache_dir = os.environ['TRITON_CACHE_DIR']
if os.path.exists(cache_dir):