mratsim

[global]
floatX = float32
device = gpu
config.openmp = True
OMP_NUM_THREADS=8
config.dnn.conv.algo_fwd='winograd'
config.dnn.conv.algo_bwd_data='winograd'
config.gcc.cxxflags="-march=native"

[lib]

mratsim

#
# /etc/makepkg.conf
#

#########################################################################
# SOURCE ACQUISITION
#########################################################################
#
#-- The download utilities that makepkg should use to acquire sources
#  Format: 'protocol::agent'

mratsim

#    yaourtrc  - Configuration for yaourt
#
#   See yaourtrc(5) for more information
#
#

# General
#AUTOSAVEBACKUPFILE=0
DEVEL=1
DEVELSRCDIR="/pkg/devel"

mratsim

# ~/.R/Makevars
PKG_LIBS = -fopenmp -lgomp
PKG_CFLAGS=-O3 -pipe -fstack-protector-strong -fopenmp -march=native

#-- Compiler and Linker Flags
# -march (or -mcpu) builds exclusively for an architecture
# -mtune optimizes for an architecture, but builds for whole processor family
CPPFLAGS=-D_FORTIFY_SOURCE=2
CFLAGS=-O3 -pipe -fstack-protector-strong -fopenmp -march=native
CXXFLAGS=-O3 -pipe -fstack-protector-strong -fopenmp -march=native

mratsim

import options
import future
import sequtils

# Unfolding the Nim Magic


## Bug https://github.com/nim-lang/Nim/issues/5647 open.
## unfold proper type signature should have Option[(T, U)] instead of Option[(T, T)]
# proc unfoldr*[T, U](f: U -> Option[(T, U)], x:U): seq[T] {. inline .}=

mratsim

template scanr[T](s: seq[T], operation: untyped): untyped =
  ## Template to scan a sequence from right to left, returning the accumulation and intermediate values.
  ## This is a foldr with intermediate steps returned

  ## @[2, 2, 3, 5].scanr(a + b) = @[48, 24, 12, 4] 
  let len = s.len

  assert len > 0, "Can't scan empty sequences"
  var result = newSeq[T](len)

mratsim

# Deleted pointer arithmetic routines
# Original data structure with pointer:

type Tensor*[B: static[Backend]; T] = object
        # Size of the datastructure is 32 bytes - perfect !
        dimensions: seq[int]
        strides: seq[int]
        offset: ptr T
        data: seq[T] # Perf note: seq are always deep copied on assignement.

mratsim

import gzip
import json
import requests
try:
    from cStringIO import StringIO
except:
    from StringIO import StringIO

# Let's fetch the Common Crawl FAQ using the CC index
resp = requests.get('http://index.commoncrawl.org/CC-MAIN-2015-27-index?url=http%3A%2F%2Fcommoncrawl.org%2Ffaqs%2F&output=json')

mratsim

type
    Context*[T] = ref object
        nodes: seq[T]

    Variable*[T; C: static[Context[T]]] = object
        tape: Context[T]
        index: int
        value: T

proc newContext*[T]: Context[T] {.noSideEffect.} =

mratsim

Achieving warp speed with Rust

Contents:

• Number one optimization tip: don't
• Never optimize blindly
• Don't bother optimizing one-time costs
• Improve your algorithms
• CPU architecture primer
• Keep as much as possible in cache
• Keep as much as possible in registers