Peter Chng pchng

## js_single_threadedness.html
<!DOCTYPE html>
<html lang="en">
  <head>
    <meta charset="utf-8">
    <title>Test of JavaScript single-threadedness</title>
  </head>
  <body>
  <h1>Test of JavaScript single-threadedness</h1>
  <button id="check">Check</button>
  <script src="https://code.jquery.com/jquery-2.1.0.min.js"></script>

## Simple Java 8 Stream API performance test
  // Test whether Stream API operations are efficiently re-ordered.
  // LOGGER is instance of org.slf4j.Logger
  @Test
  public void testStreamApiOrdering() {
    final Random rnd = new Random();
    final int nItems = 10_000;
    final List<Integer> values = new ArrayList<>(nItems);
    for (int i = 0; i < nItems; ++i) {
      values.add(rnd.nextInt(1000)); // Values between 0 and 999, inclusive.
    }

## cbc_radio_broadcast_logs.py
#!/usr/bin/env python
# Python 2.7.x
import argparse
import csv
import datetime
import sys
from collections import defaultdict

# Requires requests and beautifulsoup4: pip install beautifulsoup4 requests
import requests

## keybase.md

      
              1 file
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                pchng
                / keybase.md
            
            
              Created
              February 6, 2016 00:29
            
          
    Keybase proof

I hereby claim:

I am pchng on github.
I am pchng (https://keybase.io/pchng) on keybase.
I have a public key whose fingerprint is F722 6230 DB84 09B3 FFB2  737B B1EB 75BC 98B5 23A8

To claim this, I am signing this object:

  
## fish_prompt.fish
# Fish prompt to show username, host, CWD, Python VirtualEnv, VCS (git, svn, hg) info.
function fish_prompt --description 'Write out the prompt'
    set -l last_status $status

    if not set -q __fish_git_prompt_show_informative_status
        set -g __fish_git_prompt_show_informative_status 1
    end
    if not set -q __fish_git_prompt_hide_untrackedfiles
        set -g __fish_git_prompt_hide_untrackedfiles 1
    end

## matmul_test.cu
#include <stdio.h>

#define A 3000
#define B 4000
#define C 3000

// Computes out = left @ right, where `@` is matrix muliplication
// Dimensions:
// left:  a x b
// right: b x c

## matmul_good.cu
// Computes out = left @ right, where `@` is matrix muliplication
// Dimensions:
// left:  a x b
// right: b x c
// out:   a x c
// monolithic kernel: One thread per output element in matrix C.
// (Each thread computes the dot product between a row in `left` and a col in `right`)
__global__ void matMul(float *left, float *right, float *out, int a, int b, int c) {
  // Use y to index to rows, x to index to cols (just to match typical visualization)
  // row indexes into left, col indexes into right.

## matmul_bad.cu
// Same as above, but row/col set to x/y instead.
__global__ void matMulBad(float *left, float *right, float *out, int a, int b, int c) {
  int row = blockIdx.x * blockDim.x + threadIdx.x;
  int col = blockIdx.y * blockDim.y + threadIdx.y;

  if (row < a && col < c) {
    float sum = 0.0;
    for (int i = 0; i < b; i++) {
      // 1. If row/threadIdx.x is changing within the warp, then on each iteration the threads do a strided access:
      // They will access elements separated by a stride of b. This results in non-coalesced accesses (multiple memory reads)

## triton_simple_flashattention.py
# This is the tiling approach in: https://courses.cs.washington.edu/courses/cse599m/23sp/notes/flashattn.pdf
def flashatt_spec(q: Float32[Tensor, "200"], k: Float32[Tensor, "200"], v: Float32[Tensor, "200"]) -> Float32[Tensor, "200"]:
    x = q[:, None] * k[None, :]
    x_max = x.max(1, keepdim=True)[0]
    x = x - x_max
    x_exp = x.exp()
    soft =  x_exp  / x_exp.sum(1, keepdim=True)
    return (v[None, :] * soft).sum(1)

@triton.jit
	<!DOCTYPE html>
	<html lang="en">
	<head>
	<meta charset="utf-8">
	<title>Test of JavaScript single-threadedness</title>
	</head>
	<body>
	<h1>Test of JavaScript single-threadedness</h1>
	<button id="check">Check</button>
	<script src="https://code.jquery.com/jquery-2.1.0.min.js"></script>
	// Test whether Stream API operations are efficiently re-ordered.
	// LOGGER is instance of org.slf4j.Logger
	@Test
	public void testStreamApiOrdering() {
	final Random rnd = new Random();
	final int nItems = 10_000;
	final List<Integer> values = new ArrayList<>(nItems);
	for (int i = 0; i < nItems; ++i) {
	values.add(rnd.nextInt(1000)); // Values between 0 and 999, inclusive.
	}
	#!/usr/bin/env python
	# Python 2.7.x
	import argparse
	import csv
	import datetime
	import sys
	from collections import defaultdict

	# Requires requests and beautifulsoup4: pip install beautifulsoup4 requests
	import requests
	# Fish prompt to show username, host, CWD, Python VirtualEnv, VCS (git, svn, hg) info.
	function fish_prompt --description 'Write out the prompt'
	set -l last_status $status

	if not set -q __fish_git_prompt_show_informative_status
	set -g __fish_git_prompt_show_informative_status 1
	end
	if not set -q __fish_git_prompt_hide_untrackedfiles
	set -g __fish_git_prompt_hide_untrackedfiles 1
	end
	#include <stdio.h>

	#define A 3000
	#define B 4000
	#define C 3000

	// Computes out = left @ right, where `@` is matrix muliplication
	// Dimensions:
	// left: a x b
	// right: b x c
	// Same as above, but row/col set to x/y instead.
	__global__ void matMulBad(float left, float right, float *out, int a, int b, int c) {
	int row = blockIdx.x * blockDim.x + threadIdx.x;
	int col = blockIdx.y * blockDim.y + threadIdx.y;

	if (row < a && col < c) {
	float sum = 0.0;
	for (int i = 0; i < b; i++) {
	// 1. If row/threadIdx.x is changing within the warp, then on each iteration the threads do a strided access:
	// They will access elements separated by a stride of b. This results in non-coalesced accesses (multiple memory reads)
	# This is the tiling approach in: https://courses.cs.washington.edu/courses/cse599m/23sp/notes/flashattn.pdf
	def flashatt_spec(q: Float32[Tensor, "200"], k: Float32[Tensor, "200"], v: Float32[Tensor, "200"]) -> Float32[Tensor, "200"]:
	x = q[:, None] * k[None, :]
	x_max = x.max(1, keepdim=True)[0]
	x = x - x_max
	x_exp = x.exp()
	soft = x_exp / x_exp.sum(1, keepdim=True)
	return (v[None, :] * soft).sum(1)

	@triton.jit