tmcdonell

{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE FlexibleContexts    #-}

import Data.Array.Accelerate                       as A hiding ((!!), length)
import Data.Array.Accelerate.Debug                 as A
import Data.Array.Accelerate.Numeric.LinearAlgebra as A

import Data.Array.Accelerate.LLVM.Native           as CPU

import Control.DeepSeq

tmcdonell

{-# LANGUAGE OverloadedStrings #-}

import Text.Printf
import Foreign.Ptr
import Foreign.LibFFI -- package: libffi

import LLVM.Linking
import LLVM.OrcJIT
import LLVM.Internal.ObjectFile                    ( withObjectFile )
import qualified LLVM.Internal.OrcJIT.LinkingLayer as LL

tmcdonell

\a0 -> let a1 =
         fold
           (\x0 x1 -> x0 + x1)
           0
           (delayed
              ((intersect
                  (intersect
                     (intersect
                        (indexSlice Z :. 0 :. 0 :. All (shape a0))
                        (indexSlice Z :. 1 :. 0 :. All (shape a0)))

tmcdonell

import Data.Array.Accelerate                                        as A
import Data.Array.Accelerate.Data.Bits                              as A
import Data.Array.Accelerate.Interpreter                            as I

import Prelude                                                      as P


-- simple test

test = gather (segmentedIndices info) vec

tmcdonell

Scalar Z [1.6636479e16]
0.440618967s
1.6636479e16
0.22351678s
1.6636479e16
0.23023065s
\a0 a1 -> let a2 =
            fold
              (\x0 x1 -> x0 + x1)
              0.0

tmcdonell

a.out:     file format elf64-x86-64


Disassembly of section .init:

0000000000400408 <_init>:
  400408:       48 83 ec 08             sub    $0x8,%rsp
  40040c:       48 8b 05 e5 0b 20 00    mov    0x200be5(%rip),%rax        # 600ff8 <__gmon_start__>
  400413:       48 85 c0                test   %rax,%rax
  400416:       74 02                   je     40041a <_init+0x12>

tmcdonell

; ModuleID = 'map_1bf1a791bb25369d'
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-pc-linux-gnu"

; Function Attrs: nounwind
define void @map_1bf1a791bb25369d(i64 %ix.start, i64 %ix.end, double* noalias nocapture %out.ad0, double* noalias nocapture readonly %fv0.ad0) local_unnamed_addr #0 {
entry:
  %0 = icmp slt i64 %ix.start, %ix.end
  br i1 %0, label %while1.top.preheader, label %while1.exit

tmcdonell

; ModuleID = 'map_1bf1a791bb25369d'
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-apple-darwin15.6.0"

; Function Attrs: nounwind
define void @map_1bf1a791bb25369d(i64 %ix.start, i64 %ix.end, double* noalias nocapture %out.ad0, double* noalias nocapture readonly %fv0.ad0) local_unnamed_addr #0 {
entry:
  %0 = icmp slt i64 %ix.start, %ix.end
  br i1 %0, label %while1.top.preheader, label %while1.exit

tmcdonell

{-# LANGUAGE TemplateHaskell #-}
{-# LANGUAGE TypeOperators #-}
module ExtractSurface (extractSurface') where

import qualified Prelude as P

import Data.Word
import Data.Array.Accelerate

s333 :: Stencil3x3x3 Bool -> Exp Word8

tmcdonell

-- lib/Lib.hs
{-# LANGUAGE TemplateHaskell #-}
{-# OPTIONS_GHC -fplugin=Data.Array.Accelerate.LLVM.Native.Plugin #-}

module Lib where

import Data.Array.Accelerate              as A
import Data.Array.Accelerate.LLVM.Native  as CPU

dotp :: Vector Float -> Vector Float -> Scalar Float