Keren Zhou Jokeren

Install

git clone https://github.com/openai/triton.git;
cd triton/python;
pip install cmake; # build time dependency
pip install -e .
pip uninstall pytorch-triton -y

Expected result (-0.1250)

We couldn’t find that file to show.

Multiplexer

How many operations are still pending?
How does a thread ensure its activities are transferred before it ends? Because metrics are stored in thread local maps, it needs to get all activities and attribute them before it dies.
```
 opencl-api.c

 device_finalizer_register vs thread finalizer
```

pending_operations

	//
	// Generated by LLVM NVPTX Back-End
	//

	.version 8.0
	.target sm_80
	.address_size 64

	// .globl triton__0d1d2d3d4d56d7d89d1011d1213d1415d1617d1819d2021d2223d2425d2627d2829d3031d3233d3435d3637d3839d4041d42d
	.extern .shared .align 1 .b8 global_smem[];

	//
	// Generated by LLVM NVPTX Back-End
	//
	.version 8.0
	.target sm_80
	.address_size 64
	// .globl triton__0d1d2d3d
	.visible .entry triton__0d1d2d3d(
	.param .u64 triton__0d1d2d3d_param_0,
	.param .u64 triton__0d1d2d3d_param_1,

	#include <dlfcn.h>
	#include "tool.h"

	int main() {
	//void *handle = dlopen("./tool.so", RTLD_NOW);
	print_t func = (print_t)dlsym(RTLD_NEXT, "print");
	func();
	return 0;
	}

	#include <iostream>
	#include <chrono>

	void __attribute__ ((noinline)) init(int *arr, size_t length) {
	for (auto i = 0; i < length; ++i) {
	arr[i] = 1;
	}
	}

	class NumMatrix {
	private:
	std::vector<std::vector<int> > tree;

	void columnInit(std::vector<int> &tree_column, std::vector<int> &matrix_column) {
	size_t columns = matrix_column.size();
	for (size_t i = columns; i < tree_column.size(); ++i) {
	tree_column[i] = matrix_column[i - columns];
	}
	for (size_t i = columns - 1; i > 0; --i) {