snowolfhawk

Main inspiration comes from [here][1].

“”” Here is what a deep learning system stack would look like in nowdays.

1. Build operator level graph description language: name whatever dl frameworks you care about, and [ONNX][2]
2. Tensor primitive level graph description languages: [NNVM][3], [HLO/XLA][4], [NGraph][5]. It is close enough to the first one that you can also build graph optimization on first layer and bypass this layer.
3. DSL for description and codegen: TVM, image processing languages like [halide][6], [darkroom][7].
4. Hardcoded optimized kernel library: [nnpack][8], [cudnn][9], [libdnn][10]
5. Device dependent library: [maxas][11](assembler for NVIDIA Maxwell architecture)

snowolfhawk

#include <iostream>
#include <chrono>
#include <stdexcept>
#include <memory>
#include <fstream>
#include <iterator>
#include <algorithm>
#include <tvm/runtime/module.h>
#include <tvm/runtime/registry.h>

snowolfhawk

snowolfhawk

snowolfhawk

import tvm
import topi
import numpy as np
from tvm.testing import check_numerical_grads, estimate_performance, PerformanceEstimate
import time
import inspect
import sys
from exercise.runners import run_tvm

# Whether to dump the generated code

snowolfhawk

import tvm
from tvm import relay

def test_fuse_simple():
    """Simple testcase."""
    def before():
        x = relay.var("x", shape=(10, 20))
        y = relay.add(x, relay.const(1, "float32"))
        z = relay.exp(y)
        return relay.Function([x], z)

snowolfhawk

#include <random>
#include <iomanip>
#include <array>
#include <exception>

#define NDEBUG
#include <tvm/tvm.h>
#include <tvm/build_module.h>
#include <topi/broadcast.h>
#undef NDEBUG

snowolfhawk

snowolfhawk

"""
Quantization module for generating the calibration tables will be used by 
quantized (INT8) models from FP32 models.with bucket split,[k, k, cin, cout]
cut into "cout" buckets.
This tool is based on Caffe Framework.
"""
from __future__ import division
from __future__ import print_function
import argparse
import numpy as np

snowolfhawk

外部张量函数

虽然TVM支持透明代码生成，但有时将手动编写的代码合并到管道中也很有帮助。例如，我们想去为部分卷积和使用cuDNN和定义其他阶段。

TVM原生支持黑盒函数调用。TVM支持兼容DLPack的所有张量函数。这意味着我们可以使用POD类型（pointer，int，float）或指向DLTensor的指针作为参数调用任何函数。

from __future__ import absolute_import, print_function

import tvm