The Pochoir Project

From SuperTech Wiki

Jump to: navigation, search

Contents


Welcome to Pochoir


File:sim-space-cut.png


The Pochoir Stencil Compiler

Pochoir (pronounced "PO-shwar") is a compiler and runtime system for implementing stencil computations on multicore processors. A stencil defines the value of a grid point in a d-dimensional spatial grid at time t as a function of neighboring grid points at recent times before t. A stencil computation computes the stencil for each grid point over many time steps. Using Pochoir, a user specifies a computing kernel and boundary conditions using a simple stencil language embedded in C++. The Pochoir compiler produces cache-efficient multithreaded C++ code that can be compiled with the Intel 12.0 compiler for C++ with the Cilk multithreading extensions, which is available as part of the Intel Parallel Computer suite. The Pochoir package contains two main components: a C++ template library for debugging and testing Pochoir compliance and a domain-specific compiler written in Haskell that produces highly optimized code. (Currently, the Pochoir package has only been tested on Linux systems.)

Here is an example of a 2D heat equation coded in Pochoir:

#define mod(r,m) ((r)%(m) + ((r)<0)? (m):0)
Pochoir_Boundary_2D(heat_bv, a, t, x, y) 
  return a.get(t,mod(x,a.size(1)),mod(y,a.size(0)));
Pochoir_Boundary_End 

int main(void) {
  Pochoir_Shape_2D 2D_five_pt[] = {{1,0,0}, {0,1,0}, {0,-1,0}, {0,-1,-1}, {0,0,-1}, {0,0,1}}; 
  Pochoir_2D heat(2D_five_pt); 

  Pochoir_Array_2D(double) u(X, Y); 
  u.Register_Boundary(heat_bv); 
  heat.Register_Array(u); 

  Pochoir_Kernel_2D(heat_fn, t, x, y) 
    u(t+1, x, y) = CX * (u(t, x+1, y) - 2 * u(t, x, y) + u(t, x-1, y)) + CY * (u(t, x, y+1) - 2 * u(t, x, y) + u(t, x, y-1)) + u(t, x, y);
  Pochoir_Kernel_End
 
  for (int x = 0; x < X; ++x) 
    for (int y = 0; y < Y; ++y)    
      u(0, x, y) = rand();  

  heat.Run(T, heat_fn); 

  for (int x = 0; x < X; ++x) 
    for (int y = 0; y < Y; ++y)    
      cout << u(T, x, y); 

  return 0;
}

Pochoir is an open-source software project hosted by Supertech group at CSAIL, MIT. You are invited to contribute in any form (documentation, translation, writing code, fixing bugs, porting to other platforms, etc.). We are particularly interested in developing a set of real-world example stencil computations.


Performance Graphs

Comparing the number of grid points processed per second (semilogarithmic scale) for Pochoir-generated code on 12 cores versus serial- and parallel-loop implementations. In all figures, the top curve is for the Pochoir-generated code, the middle curve is for parallel loops, and the bottom curve is for serial loops. (a) A 3D wave equation with a nonperiodic boundary condition executing for 1000 time steps. (b) A 2D heat equation on a torus executing for 3200 time steps. (c) 1D pairwise sequence alignment (no time step). (d) A Lattice Boltzmann method with a nonperiodic boundary condition for 3000 time steps.


Pochoir Team

The Pochoir project is joint research by the Supertech Research Group at MIT, Intel Corporation, Department of Computer Science, State University of New York at Stony Brook, Nanostructures and Computation Group at MIT, and Fudan University in China. The core team includes the following researchers:


Current Release

Pochoir 0.5 [1] is now released (Feb, 2011)

Please send your name and affiliation to Pochoir Development List to let us know that you are interested in using Pochoir, so that we can provide better technical support to all our known customers, which includes bug fix, new release notification, new feature incorporation, etc.

Pochoir is released under the GNU Public License version 3.0.


Documentation

The Pochoir Stencil Compiler, Yuan Tang, Rezaul Alam Chowdhury, Bradley C. Kuszmaul, Chi-Keung Luk, and Charles E. Leiserson, to appear in 23rd ACM Symposium on Parallelism in Algorithms and Architectures (SPAA'2011), 2011.

Coding Stencil Computations Using the Pochoir Stencil-Specification Language, Yuan Tang, Rezaul Alam Chowdhury, Chi-Keung Luk, Charles E. Leiserson, to appear in 3rd USENIX Workshop on Hot Topics in Parallelism (HotPar'11), 2011.

Manual

Acknowledgment

The Pochoir project is supported in part by a grant from Intel Corporation and in part by NSF Grants CCF-0937860 and CNS-1017058.



Copyright  ©  2010-2011 Yuan Tang and Charles E. Leiserson, All Rights Reserved.

Personal tools
Members Only
Off Topic