bwasti

# split

for i in range(107):
  a[i] = b[i] + c[i]

for i0 in range(10):
  for i1 in range(10):
    a[i0 * 10 + i1] = b[i0 * 10 + i1] + c[i0 * 10 + i1]
for i0 in range(7):
  a[100 + i0] = b[100 + i0] + c[100 + i0]

bwasti

template<bool...> struct bool_pack;
template<bool... bs> 
using all_true = std::is_same<bool_pack<bs..., true>, bool_pack<true, bs...>>;

template<class R, class... Ts>
using are_all_constructible = all_true<std::is_constructible<R, Ts>::value...>;

template<typename... Ts>
struct ivalue_constructible_tuple {
  constexpr static bool value = are_all_constructible<c10::IValue, Ts...>::value;

bwasti

bwasti

>>> import this
The Zen of Python, by Tim Peters

Beautiful is better than ugly.
Explicit is better than implicit.
Simple is better than complex.
Complex is better than complicated.
Flat is better than nested.
Sparse is better than dense.
Readability counts.

bwasti

/*
How to run:

PT_DIR=$(python -c 'import os, torch; print(os.path.dirname(os.path.realpath(torch.__file__)))')
g++ -O3 test.cc -o test -I$PT_DIR/include -L$PT_DIR/lib -ltorch -lc10
for((i=1;i<=100000;i*=2)); do LD_LIBRARY_PATH="$PT_DIR/lib:$LD_LIBRARY_PATH" ./test $i; done

*/
#include <ATen/core/Dict.h>
#include <chrono>

bwasti

Size 1
                       DictPtr : 335 usec
            std::unordered_map : 442 usec
std::unordered_map + std::move : 381 usec
Size 2
                       DictPtr : 347 usec
            std::unordered_map : 553 usec
std::unordered_map + std::move : 459 usec
Size 4
                       DictPtr : 553 usec

bwasti

PyTorch now has an official TVM-based backend, torch_tvm. Usage is simple:

import torch_tvm
torch_tvm.enable()

That's it! PyTorch will then attempt to convert all operators it can to known Relay operators during its JIT compilation process.

Background

bwasti

PyTorch now has an official TVM-based backend, [torch_tvm](https://github.com/pytorch/tvm).  Usage is simple:

```
import torch_tvm
torch_tvm.enable()
```

That's it!  PyTorch will then attempt to convert all operators it can to known Relay operators during its JIT compilation process.

### Background

bwasti

diff --git a/include/tvm/runtime/c_runtime_api.h b/include/tvm/runtime/c_runtime_api.h
index 1a1a8da6..2e15b179 100644
--- a/include/tvm/runtime/c_runtime_api.h
+++ b/include/tvm/runtime/c_runtime_api.h
@@ -85,6 +85,7 @@ typedef enum {
   kStr = 11U,
   kBytes = 12U,
   kNDArrayContainer = 13U,
+  kManagedArrayHandle = 14U,
   // Extension codes for other frameworks to integrate TVM PackedFunc.

bwasti

import tvm
from tvm import relay
import torch
import torch.nn.functional as F
import inspect
import ast
import numpy as np

_parsed_functions = dict()