malfet’s gists

malfet / simple-kineto.py

Created March 3, 2021 20:37

	#!/usr/bin/env python

	import torch
	from torch.autograd.profiler import profile as _profile

	def workload():
	s1 = torch.cuda.Stream(device="cuda")
	s2 = torch.cuda.Stream(device="cuda")

	with torch.cuda.stream(s1):

malfet / cifar10_benchmark.py

Created April 5, 2021 14:36

Run CIFAR

	#!/usr/bin/env python3
	# Results of recent runs:
	# Mac Apple M1 50.3 sec
	# Mac Intel(R) Core(TM) i9-9980HK CPU @ 2.40GH in 61.1 sec
	# Linux Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz in 53.5 sec
	import time
	import torch
	import torch.nn as nn
	import torch.nn.functional as F
	import torch.optim as optim

malfet / gist:23e71fbb909707166bbb8cf24de8b17e

Created May 26, 2021 22:40

Triggering new pipeline on Circle and checking it's status

	% curl --request POST --url https://circleci.com/api/v2/project/gh/pytorch/pytorch/pipeline --data '{"branch":"pull/59020/head", "parameters": {"run_slow_gradcheck_build": true}}' --header 'content-type: application/json' --header 'Circle-Token: XXXXXX'
	{
	"number" : 328134,
	"state" : "pending",
	"id" : "d23f0239-1cd0-4d22-8965-d40f8c7bbd73",
	"created_at" : "2021-05-26T22:37:55.955Z"
	}
	% curl https://circleci.com/api/v2/project/gh/pytorch/pytorch/pipeline/328134
	{
	"id" : "d23f0239-1cd0-4d22-8965-d40f8c7bbd73",

malfet / cudnn-xmma.cpp

Last active October 26, 2022 06:52

	// If linked against cudnn dynamically, i.e. using following command:
	// g++ cudnn-xmma.cpp -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudnn -lcudart
	// Then nvprof shows following kernels executed on RTX2080:
	// Type Time(%) Time Calls Avg Min Max Name
	// GPU activities: 64.57% 20.416us 1 20.416us 20.416us 20.416us void xmma_cudnn::gemm::kernel<xmma_cudnn::implicit_gemm::fprop::Kernel_traits<xmma_cudnn::Turing_hmma_fp32_traits, xmma_cudnn::Cta_tile<xmma_cudnn::Turing, int=64, int=32, int=64, int=2, int=1, int=2, int=1, int=1>, xmma_cudnn::implicit_gemm::fprop::Gmem_tile_a_t<xmma_cudnn::Turing_hmma_fp32_traits, xmma_cudnn::Cta_tile<xmma_cudnn::Turing, int=64, int=32, int=64, int=2, int=1, int=2, int=1, int=1>, xmma_cudnn::implicit_gemm::Input_related<int=0, int=0, int=0, bool=0>, int=16, bool=0, xmma_cudnn::implicit_gemm::fprop::Gmem_tile_base_a<xmma_cudnn::Turing_hmma_fp32_traits, xmma_cudnn::Cta_tile<xmma_cudnn::Turing, int=64, int=32, int=64

malfet / neon-trunc.cpp

Created June 11, 2021 14:23

Implements trunc using vcvtq vs vrndq

	#include <arm_neon.h>
	#include <iostream>

	int main(void) {
	const float inp[4] = { 3.1, 2.7, -2.9, -1.0000000200408773e+20 };
	std::cout << "inp= " << inp[0] << " " << inp[1] << " " << inp[2] << " " << inp[3] << std::endl;
	float32x4_t f = vld1q_f32(inp);
	float32x4_t old_trunc = vcvtq_f32_s32(vcvtq_s32_f32(f));
	float32x4_t new_trunc = vrndq_f32(f);
	float out[4];

malfet / conda_remove_abi.py

Created August 5, 2021 04:46

Ugly script that removes python_abi dep from conda packages

	#!/usr/bin/env python3
	from conda_build.convert import extract_temporary_directory, create_target_archive
	import shutil
	import os
	import sys

	def remove_line_with_pattern(fname, pattern):
	with open(fname) as f:
	lines = f.read().split('\n')
	del_idx = []

malfet / mup-smaller.cpp

Created September 28, 2021 13:49

clang-13 crash using omp critical pragma

	// c++ -std=c++14 -O2 -fPIC -Xpreprocessor -fopenmp mup-smaller.cpp
	//
	#include <omp.h>
	#include <atomic>
	#include <exception>
	#include <algorithm>

	#define C10_UNLIKELY(expr) (__builtin_expect(static_cast<bool>(expr), 0))

	namespace at {

malfet / whatsmyname.cpp

Created October 7, 2021 16:41

Prints name of the template argument

	#include <iostream>
	#include <string>

	template<typename T>
	struct WhatsMyName {
	WhatsMyName() {
	std::cout << __PRETTY_FUNCTION__ << std::endl;
	}
	};

malfet / gist:175d76815eebf8b2f19005cb80105b45

Created November 1, 2021 17:44

pytorch.github.io broken links

	URL `resources/fonts/dejavu.css'
	Parent URL file:///home/nshulga/git/pytorch.github.io/javadoc/stylesheet.css, line 3, col 12
	Real URL file:///home/nshulga/git/pytorch.github.io/javadoc/resources/fonts/dejavu.css
	Result Error: URLError: <urlopen error [Errno 2] No such file or directory: '/home/nshulga/git/pytorch.github.io/javadoc/resources/fonts/dejavu.css'>

	URL `../fonts/FreightSans/freight-sans-bold.woff'
	Parent URL file:///home/nshulga/git/pytorch.github.io/assets/main.css, line 1, col 147264
	Real URL file:///home/nshulga/git/pytorch.github.io/fonts/FreightSans/freight-sans-bold.woff
	Result Error: URLError: <urlopen error [Errno 2] No such file or directory: '/home/nshulga/git/pytorch.github.io/fonts/FreightSans/freight-sans-bold.woff'>

malfet / hello.cu

Last active November 24, 2021 21:05

CUDA Hello world

	#include <stdio.h>
	__global__ void kernel() {
	printf("Hello World of CUDA %d\n", threadIdx.x);
	}

	int main() {
	kernel<<<1,1>>>();
	return cudaDeviceSynchronize();
	}

Nikita Shulga malfet