Difference between revisions of "GPUSquad"

From CDOT Wiki
Jump to: navigation, search
(→‎Idea 1 - Jacobi Method for 2D Poisson Problem)
m (→‎Team Members)
 
(37 intermediate revisions by 2 users not shown)
Line 13: Line 13:
 
</source>
 
</source>
 
== Team Members ==  
 
== Team Members ==  
−
# [mailto:tsarkar3@myseneca.ca?subject=dps915 Tanvir Sarkar]
+
# TS
 
# [mailto:moverall@myseneca.ca?subject=dps915 Michael Overall]
 
# [mailto:moverall@myseneca.ca?subject=dps915 Michael Overall]
 
# [mailto:ikrasnyanskiy@myseneca.ca?subject=gpu610 Igor Krasnyanskiy]
 
# [mailto:ikrasnyanskiy@myseneca.ca?subject=gpu610 Igor Krasnyanskiy]
−
# [mailto:tsarkar3@myseneca.ca;moverall@myseneca.ca;ikrasnyanskiy@myseneca.ca?subject=dps915gpu610 Email All]
+
# [mailto:moverall@myseneca.ca;ikrasnyanskiy@myseneca.ca?subject=dps915gpu610 Email All]
  
 
== Progress ==
 
== Progress ==
Line 243: Line 243:
 
PNG image should be created in the same folder as the jacobi out.# files
 
PNG image should be created in the same folder as the jacobi out.# files
 
</source>
 
</source>
 +
 +
The images look something like this:
 +
m = 32, n = 1024
 +
 +
[[File:dps915_gpusquad_jacobi_output.png]]
  
 
<nowiki>************************</nowiki>
 
<nowiki>************************</nowiki>
Line 293: Line 298:
  
 
</source>
 
</source>
−
 
+
<nowiki>************************</nowiki>
 +
<pre style="color: green">
 
The hotspot seems to be the double for-loop based on m and n in the Jacobi iterations code of the dojacobi() function. I believe these matrix calculations could be parallelized for improved performance. Note that the for-loop that the double loop is inside of is based on a constant numbers, iters, so it doesn't grow with the problem size. It would be O(iters * n^2) which is still O(n^2) not O(n^3).
 
The hotspot seems to be the double for-loop based on m and n in the Jacobi iterations code of the dojacobi() function. I believe these matrix calculations could be parallelized for improved performance. Note that the for-loop that the double loop is inside of is based on a constant numbers, iters, so it doesn't grow with the problem size. It would be O(iters * n^2) which is still O(n^2) not O(n^3).
 +
</pre>
  
 
==== Idea 2 - LZW Compression ====
 
==== Idea 2 - LZW Compression ====
Line 520: Line 527:
 
POTENTIAL FOR PARALLELIZATION:
 
POTENTIAL FOR PARALLELIZATION:
  
−
The compress() function performs similar operations on a collection of text, however it relies on a dictionary and an expanding string to be tokenized in a dictrionary. This could potentially be paralellized through a divide and conquer strategy where gpu blocks with shared caches share their own dictionary and iterate over their own block of text.
+
The compress() function performs similar operations on a collection of text, however it relies on a dictionary and an expanding string to be tokenized in a dictrionary. This could potentially be paralellized through a divide and conquer strategy where gpu blocks with shared caches share their own dictionaries and iterate over their own block of text. This, however would not be particularly useful because LZW compression relies on a globally accessible dictionary that can be continuously modified. Having multiple threads try to access the same dictionary tokens in global memory would create race conditions, and creating block-local dictionaries in shared memory would reduce the efficacy of having a large globally available dictionary to tokenize strings.
 +
 
 +
What all of this means is that LZW compression would probably be a poor candidate for parallelization due to the way that parallel memory would try to update the token dictionary.
  
 
==== Idea 3 - MergeSort ====
 
==== Idea 3 - MergeSort ====
Line 659: Line 668:
  
 
=== Assignment 2 ===
 
=== Assignment 2 ===
 +
 +
We chose to move forward with the Jacobi Method for 2D Poisson Equations code.
 +
 +
We parallelized the original code by placing the jacobi calculations into a kernel. For this initial parallel version, we only used 1D threading and had each thread run a for loop for the other dimension.
 +
 +
The iters loop launches a kernel for each iteration and we use double buffering (where we choose to launch the kernel with either d_a, d_b or d_b, d_a) since we can't simply swap pointers like in the serial code.
 +
 +
Double buffering is needed so that there is a static version of the the matrix before the current set of calculations are done. Since an element of the matrix is calculated based on elements on all four sides, if one of those elements updated itself while calculations were being done on another element, you would end up with the wrong answer/race conditions.
 +
 +
<nowiki>****************</nowiki>
 +
 +
CODE
 +
 +
<nowiki>****************</nowiki>
 +
 
<source>
 
<source>
  
Line 675: Line 699:
 
using namespace std;
 
using namespace std;
  
 +
</source>
 +
<pre style="color: blue">
 
__global__ void matrixKernel(float* a, float* b, float dcent, int n, int m, int xmin, int xmax,
 
__global__ void matrixKernel(float* a, float* b, float dcent, int n, int m, int xmin, int xmax,
 
int ymin, int ymax, float dx, float dy, float dxxinv, float dyyinv) {//MODIFY to suit algorithm
 
int ymin, int ymax, float dx, float dy, float dxxinv, float dyyinv) {//MODIFY to suit algorithm
  
−
int j = threadIdx.x + 1;
+
int j = blockDim.x * blockIdx.x + threadIdx.x + 1;
 +
 
 
//above: we are using block and thread indexes to replace some of the iteration logic
 
//above: we are using block and thread indexes to replace some of the iteration logic
 
if (j < n - 1) {
 
if (j < n - 1) {
Line 694: Line 721:
  
 
}
 
}
 +
</pre>
 +
<source>
 
// Set grid size and number of iterations
 
// Set grid size and number of iterations
 
const int save_iters = 20;
 
const int save_iters = 20;
Line 781: Line 810:
 
dim3 dGrid(nblocks);
 
dim3 dGrid(nblocks);
 
dim3 dBlock(1024);
 
dim3 dBlock(1024);
−
 
+
</source>
 +
<pre style="color: blue">
 
// Carry out Jacobi iterations
 
// Carry out Jacobi iterations
 
for (k = 1; k <= total_iters; k++) {
 
for (k = 1; k <= total_iters; k++) {
Line 806: Line 836:
 
}
 
}
 
}
 
}
 +
</pre>
 +
<source>
 
cudaMemcpy(a, d_a, n* m * sizeof(float), cudaMemcpyDeviceToHost);
 
cudaMemcpy(a, d_a, n* m * sizeof(float), cudaMemcpyDeviceToHost);
 
outputImage("jacobi out", a);
 
outputImage("jacobi out", a);
Line 826: Line 858:
 
return 0;
 
return 0;
 
}
 
}
 +
</source>
 +
 +
<nowiki>****************</nowiki>
 +
 +
A2 TIMINGS
  
 +
<nowiki>****************</nowiki>
 +
 +
[[File:dps915_gpusquad_a2_chart.png]]
 +
 +
=== Assignment 3 ===
 +
Optimization techniques attempted
 +
* Get rid of the for loop in the kernel and use 2D threading within blocks
 +
* Use gpu constant memory for jacobi calculation constants
 +
* Utilize the ghost cell pattern for shared memory within blocks
 +
 +
The ghost cell pattern is a technique used to allow threads in a particular block to access data that would normally be inside another block.
 +
We make a shared memory array larger than the number of threads per block in a particular dimension, then when we are at the first or last thread,
 +
can copy a neighbouring cells' data based on a global index, to our local ghost cell.
 +
 +
The following info and images are taken from: http://people.csail.mit.edu/fred/ghost_cell.pdf
 +
 +
Ghost Cell Pattern Abstract:
 +
<pre>
 +
Many problems consist of a structured grid of points that
 +
are updated repeatedly based on the values of a fixed set
 +
of neighboring points in the same grid. To parallelize these
 +
problems we can geometrically divide the grid into chunks
 +
that are processed by different processors. One challenge
 +
with this approach is that the update of points at the periphery
 +
of a chunk requires values from neighboring chunks.
 +
These are often located in remote memory belonging to different
 +
processes. The naive implementation results in a lot
 +
of time spent on communication leaving less time for useful
 +
computation. By using the Ghost Cell Pattern communication
 +
overhead can be reduced. This results in faster time to
 +
completion.
 +
</pre>
 +
 +
In our code, our main jacobi calculations look like this (when using global indexing):
 +
<source>
 +
a[ij] = (input + dxxinv*(b[ij - 1] + b[ij + 1])
 +
+ dyyinv*(b[ij - m] + b[ij + m]))*dcent;
 
</source>
 
</source>
 +
* b[ij - 1] is a cell's left neighbour
 +
* b[ij + 1] is a cell's right neighbour
 +
* b[ij - m] is a cell's top neighbour (we represent a 2D array as a 1D array, so subtract row length to go one cell up)
 +
* b[ij + m] is a cell's bottom neighbour (add row to go one cell down)
  
−
=== Assignment 3 ===
+
[[File:Dps_915_gpusquad_needneighbour.png]]
 +
 
 +
[[File:Dps_915_gpusquad_ghostcellexample.png]]
 +
 
 +
We use a 1D grid of blocks which have 2D threads.
 +
Thus the first block needs ghost cells on the right side, the last block needs ghost cells on the left side, and
 +
all other blocks in the middle need ghost cells on both the left and right sides.
 +
 
 +
In A2 and A3 we only scaled the columns (we would have, had we not mixed up the indexing--the 2d version scales along the n dimension, while the 1d versions scale properly along the n dimension), and the other dimension always stayed at 32. This made it easy for us to keep the grid of blocks one dimensional.
 +
If we scaled the rows as well as the columns we would most likely have need to use a two dimensional grid of blocks.
 +
This would make the problem more difficult as we would then not only need ghost columns, but also ghost rows on the top and bottom of blocks. This would likely cause synchronization problems because we would need to make sure that block execution was synchronized along 2 dimensions instead of just one.
 +
 
 +
[[File:dps915_gpusquad_2Dghostcells.png]]
 +
 
 +
<nowiki>****************</nowiki>
 +
 
 +
CODE FOR 1D GRID WITH 2D THREAD SETUP (kernel contains shared and global memory versions)
 +
 
 +
<nowiki>****************</nowiki>
 +
 
 +
<source>
 +
#include <cuda.h>
 +
#include "cuda_runtime.h"
 +
#include <device_launch_parameters.h>
 +
#include <stdio.h>
 +
#include <device_functions.h>
 +
#include <cuda_runtime_api.h>
 +
#ifndef __CUDACC__
 +
#define __CUDACC__
 +
#endif
 +
// Load standard libraries
 +
#include <cstdio>
 +
#include <cstdlib>
 +
#include <iostream>
 +
#include <fstream>
 +
#include <cmath>
 +
#include <chrono>
 +
using namespace std;
 +
const int ntpb = 1024;//TOTAL number of threads per block
 +
const int ntpbXY = 32;
 +
__device__ __constant__ int d_m;
 +
__device__ __constant__ int d_n;
 +
__device__ __constant__ float d_xmin;
 +
__device__ __constant__ float d_xmax;
 +
__device__ __constant__ float d_ymin;
 +
__device__ __constant__ float d_ymax;
 +
__device__ __constant__ float d_pi;
 +
__device__ __constant__ float d_omega;
 +
__device__ __constant__ float d_dx;
 +
__device__ __constant__ float d_dy;
 +
__device__ __constant__ float d_dxxinv;
 +
__device__ __constant__ float d_dyyinv;
 +
__device__ __constant__ float d_dcent;
 +
__device__ __constant__ int d_nBlocks;
 +
__device__ __constant__ int d_ntpbXY;
 +
void check(cudaError_t err) {
 +
  if (err != cudaSuccess)
 +
    std::cerr << "ERROR: *** " << cudaGetErrorString(err) << " ****" << std::endl;
 +
}
 +
 
 +
</source>
 +
<pre style="color: blue">
 +
//every thread does a single calculation--remove iterations
 +
//-take everything from one dimensional global, and put into 2d shared--shared has ghost b/c need for threads--
 +
//--global is 1d indexed ,shared is 2d indexed.
 +
//TODO:  remove for loop from kernel and index entirely with blockIdx, blockDim, and threadIdx (check workshop 7 for indexing by shared memory within blocks)
 +
__global__ void matrixKernel(float* a, float* b) {//MODIFY to suit algorithm
 +
    //int j= blockDim.x*blockIdx.x + threadIdx.x+1;//TODO: CHECK THIS DOESN'T MESS UP GLOBAL EXECUTION
 +
  int j = threadIdx.y;
 +
  int i = blockDim.x*blockIdx.x + threadIdx.x;
 +
    //int i = threadIdx.y + 1;//original code has k = 1 and <=total_iters--since k will now start at 0, set end at < total_iters
 +
                                                      //above: we are using block and thread indexes to replace some of the iteration logic
 +
  int tj = threadIdx.y;
 +
  int ti = threadIdx.x;
 +
  int tij = threadIdx.x + ntpbXY*threadIdx.y;
 +
 
 +
   
 +
  int  ij = i + d_m*j;
 +
 
 +
  //GLOBAL MEMORY VERSION (COMMENT OUT SHARED SECTION BELOW TO RUN):==================================
 +
//  if (!(blockIdx.x == 0 && threadIdx.x == 0 || blockIdx.x == (d_nBlocks - 1) && threadIdx.x == (d_ntpbXY - 1)) && threadIdx.y != 0 && threadIdx.y != (d_ntpbXY - 1)) {
 +
//  float x = d_xmin + i*d_dx, y = d_ymin + j*d_dy;
 +
  //  float input = abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
 +
  // a[ij] = (input + d_dxxinv*(b[ij - 1] + b[ij + 1])
 +
    //  + d_dyyinv*(b[ij - d_m] + b[ij + d_m]))*d_dcent;
 +
//  }
 +
 
 +
 
 +
//SHARED MEMORY SETUP (COMMENT OUT FOR GLOBAL VERSION)
 +
  __shared__ float bShared[ntpbXY+2][ntpbXY];
 +
  //NOTES ON SHARED INDEX OFFSET:
 +
  //x offset is -1, so to get the current index for the thread for global, it is bShared[ti+i][tj]
 +
  //-to get global x index -1 is bShared[ti][tj]
 +
  //-to get global x index is bShared[ti+1][tj]
 +
  //-to get global x index +1 is bShared[ti+2][tj]
 +
 
 +
  //set left ghost cells
 +
  if (threadIdx.x == 0 && blockIdx.x != 0)
 +
    bShared[ti][tj] = b[ij - 1];
 +
 
 +
  //set right ghost cells
 +
  if (threadIdx.x == ntpbXY - 1 && blockIdx.x != d_nBlocks - 1)
 +
    bShared[ti+2][tj] = b[ij+1];
 +
 
 +
  //set ghost cell for current thread relative to global memory
 +
  bShared[ti+1][tj] = b[d_m* j + i];
 +
 
 +
  __syncthreads();
 +
 
 +
 +
  //SHARED MEMORY VERSION:==========
 +
  if (!(blockIdx.x == 0 && threadIdx.x == 0 || blockIdx.x == (d_nBlocks - 1) && threadIdx.x == (d_ntpbXY - 1)) && threadIdx.y != 0 && threadIdx.y != (d_ntpbXY - 1)) {
 +
    float x = d_xmin + i*d_dx, y = d_ymin + j*d_dy;
 +
    float input = abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
 +
        a[d_m* j + i] = (input + d_dxxinv*(bShared[ti][tj] + bShared[ti+2][tj])//TODO: does the program logic allow for this to go out of range??
 +
    + d_dyyinv*(bShared[ti+1][tj-1] + bShared[ti+1][tj+1]))*d_dcent;
 +
      __syncthreads();
 +
   
 +
    }
 +
}
 +
</pre>
 +
<source>
 +
// Set grid size and number of iterations
 +
const int save_iters = 20;
 +
const int total_iters = 5000;
 +
const int error_every = 2;
 +
//HERE:
 +
const int m = 32, n =32 ;
 +
const float xmin = -1, xmax = 1;
 +
const float ymin = -1, ymax = 1;
 +
// Compute useful constants
 +
const float pi = 3.1415926535897932384626433832795;
 +
const float omega = 2 / (1 + sin(2 * pi / n));
 +
const float dx = (xmax - xmin) / (m - 1);//TODO: modify to fit kernel?
 +
const float dy = (ymax - ymin) / (n - 1);//TODO: modify to fit kernel?
 +
const float dxxinv = 1 / (dx*dx);
 +
const float dyyinv = 1 / (dy*dy);
 +
const float dcent = 1 / (2 * (dxxinv + dyyinv));
 +
// Input function
 +
inline float f(int i, int j) {
 +
    float x = xmin + i*dx, y = ymin + j*dy;
 +
    return abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
 +
}
 +
// Common output and error routine
 +
void outputImage(char* filename, float *a) {
 +
    // Computes the error if sn%error every==0
 +
    // Saves the matrix if sn<=save iters
 +
    int i, j, ij = 0, ds = sizeof(float);
 +
    float x, y, data_float; const char *pfloat;
 +
    pfloat = (const char*)&data_float;
 +
    ofstream outfile;
 +
    static char fname[256];
 +
    sprintf(fname, "%s.%d", filename, 101);
 +
    outfile.open(fname, fstream::out
 +
        | fstream::trunc | fstream::binary);
 +
    data_float = m; outfile.write(pfloat, ds);
 +
    for (i = 0; i < m; i++) {
 +
        x = xmin + i*dx;
 +
        data_float = x; outfile.write(pfloat, ds);
 +
    }
 +
    for (j = 0; j < n; j++) {
 +
        y = ymin + j*dy;
 +
        data_float = y;
 +
        outfile.write(pfloat, ds);
 +
        for (i = 0; i < m; i++) {
 +
            data_float = a[ij++];
 +
            outfile.write(pfloat, ds);
 +
        }
 +
    }
 +
    outfile.close();
 +
}
 +
void dojacobi() {
 +
    int i, j, ij, k;
 +
    float error, z;
 +
    float *a, *b;
 +
    float *u;
 +
    float *v;
 +
    u = new float[m*n];
 +
    v = new float[m*n];
 +
    // Set initial guess to be identically zero
 +
    for (ij = 0; ij < m*n; ij++) u[ij] = v[ij] = 0;
 +
    a = v; b = u;
 +
    float* d_a;
 +
    float* d_b;
 +
    int nBlocks = ((n*m + ntpb - 1) / ntpb);
 +
    //malloc
 +
    cudaMalloc((void**)&d_a, m*n * sizeof(float));
 +
    cudaMalloc((void**)&d_b, m*n * sizeof(float));
 +
    cudaMemcpy(d_a, a, n* m * sizeof(float), cudaMemcpyHostToDevice);
 +
    cudaMemcpy(d_b, b, n* m * sizeof(float), cudaMemcpyHostToDevice);
 +
    dim3 dGrid(nBlocks, 1);//1D grid of blocks (means we don't have to ghost 1st and last rows of matrices w/i a block
 +
    dim3 dBlock(32, 32);
 +
  cudaMemcpyToSymbol(d_nBlocks, &nBlocks, sizeof(int));
 +
  cudaMemcpyToSymbol(d_ntpbXY, &ntpbXY, sizeof(int));
 +
    // Carry out Jacobi iterations
 +
    for (k = 1; k <= total_iters; k++) {
 +
        if (k % 2 == 0) {
 +
            cudaError_t error = cudaGetLastError();
 +
//RUN KERNEL:
 +
            matrixKernel << <dGrid, dBlock >> > (d_a, d_b);
 +
            cudaDeviceSynchronize();
 +
      check(cudaGetLastError());
 +
        }
 +
        else {
 +
            cudaError_t error = cudaGetLastError();
 +
            matrixKernel << <dGrid, dBlock >> > (d_b, d_a);
 +
            cudaDeviceSynchronize();
 +
            if (cudaGetLastError()) {
 +
                std::cout << "error";
 +
            }
 +
        }
 +
    }
 +
    cudaMemcpy(a, d_a, n* m * sizeof(float), cudaMemcpyDeviceToHost);
 +
    outputImage("jacobi out", a);
 +
    cudaFree(d_a);
 +
    cudaFree(d_b);
 +
    delete[] u;
 +
    delete[] v;
 +
}
 +
int main() {
 +
    cudaError_t cuerr;
 +
   
 +
    cuerr = cudaMemcpyToSymbol(d_m, &m, sizeof(int));
 +
    if (cudaGetLastError()) {
 +
        std::cout << cudaGetErrorString(cuerr) << std::endl;
 +
    }
 +
</source>
 +
<pre style="color: blue">
 +
    cudaMemcpyToSymbol(d_n, &n, sizeof(int));
 +
    cudaMemcpyToSymbol(d_xmin, &xmin, sizeof(float));
 +
    cudaMemcpyToSymbol(d_xmax, &xmax, sizeof(float));
 +
    cudaMemcpyToSymbol(d_ymin, &ymin, sizeof(float));
 +
    cudaMemcpyToSymbol(d_ymax, &ymax, sizeof(float));
 +
    cudaMemcpyToSymbol(d_pi, &pi, sizeof(float));
 +
    cudaMemcpyToSymbol(d_omega, &pi, sizeof(float));
 +
    cudaMemcpyToSymbol(d_dx, &dx, sizeof(float));
 +
    cudaMemcpyToSymbol(d_dy, &dy, sizeof(float));
 +
</pre>
 +
<source>
 +
    cuerr = cudaMemcpyToSymbol(d_dxxinv, &dxxinv, sizeof(float));
 +
    if (cudaGetLastError()) {
 +
        std::cout << "error" << std::endl;
 +
        std::cout << cudaGetErrorString(cuerr) << std::endl;
 +
    }
 +
    cudaMemcpyToSymbol(d_dyyinv, &dyyinv, sizeof(float));
 +
    cuerr = cudaMemcpyToSymbol(d_dcent, &dcent, sizeof(float));
 +
    if (cudaGetLastError()) {
 +
        //std::cout << "error";
 +
        std::cout << cudaGetErrorString(cuerr) << std::endl;
 +
    }
 +
    std::chrono::steady_clock::time_point ts, te;
 +
    ts = std::chrono::steady_clock::now();
 +
    dojacobi();
 +
    te = std::chrono::steady_clock::now();
 +
    std::chrono::steady_clock::duration duration = te - ts;
 +
    auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);
 +
    std::cout << "Parallel Code Time: " << ms.count() << " ms" << std::endl;
 +
    cudaDeviceReset();
 +
    return 0;
 +
}
 +
</source>
 +
 
 +
<nowiki>****************</nowiki>
 +
 
 +
A NOTE ON SCALABILITY:
 +
 
 +
In our attempts to make the kernel scalable with ghost cells, we scaled along one dimension. However, we were inconsistent in our scaling. The 1D kernel scaled along the n (y) dimension while the 2d kernels scaled along the m (x) dimension. Scaling along the x dimension, while allowing results to be testable between serial and 2D parallelized versions of the code, produced distributions that were strangely banded and skewed. In other words, we made the code render weird things faster:
 +
 
 +
[[File:MDimensionScale.png]]
 +
 
 +
FINAL TIMINGS <pre style="color: red"> THE GRAPH IMMEDIATELY BELOW IS INCORRECT: there was an error recording the 1D runtimes for assignment 2</pre>
 +
 
 +
<nowiki>****************</nowiki>
 +
 
 +
[[File:dps915_gpusquad_a3chart.png]]
 +
 
 +
<nowiki>****************</nowiki>
 +
 
 +
PROPER TIMINGS:
 +
 
 +
[[File:Code_timings2.png]]
 +
 
 +
Blue = Serial (A1)
 +
Orange = Parallel (A2)
 +
Grey = Optimized Global (A3)
 +
Yellow = Optimized Shared (A3)
 +
 
 +
The above graph includes the total run times for the serial code, the 1D kernel from assignment 2, a kernel with global and constant memory with a 2D thread arrangement, and the same 2D arrangement but with shared memory utilizing ghost cells.
 +
 
 +
We found that the most efficient version of the code was the 2d version that used constant memory and did not use shared memory. Because the shared memory version of the kernel required synchronization of threads to allocate shared memory every time a kernel was run, and a kernel was run 5000 times for each version of our code, this increased overhead for memory setup actually made the execution slower than the version with global memory.
 +
 
 +
We found that the most efficient version of the code was the 2d version that used constant memory and did not use shared memory. Because the shared memory version of the kernel required synchronization of threads to allocate shared memory every time a kernel was run, and a kernel was run 5000 times for each version of our code, the if statements required to set up the ghost cells for shared memory may have created a certain amount of warp divergence, thus slowing down the runtimes of each individual kernel.
 +
 
 +
Below, are two images that show 4 consecutive kernel runs for both global and shared versions of the code. It is apparent that shared kernel runs actually take more time than the global memory versions.
 +
 
 +
 
 +
TIMES FOR THE GLOBAL KERNEL
 +
[[File:kernelGlobalTimes.png]]
 +
 
 +
 
 +
TIMES FOR THE SHARED KERNEL
 +
 
 +
[[File:sharedKernelTimes.png]]
 +
 
 +
Note how the run times for each kernel with shared memory are significantly longer than those with global.
 +
 
 +
To try to determine if this issue was one of warp divergence, we tried to time a kernel with global memory that also initialized shared memory, although referenced global memory when carrying out the actual calculations:
 +
 
 +
[[File:GlobalInitSharedKernelTimes.png]]
 +
 
 +
The run of a kernel that allocated shared memory using a series of if statements, but executed instructions using global memory is shown in the figure above. While slightly longer than the run with global memory where shared memory is not initialized for ghost cells, it still takes less time to run than the version with Global memory. It is likely that Our group's attempts to employ shared memory failed because we did not adequately schedule or partition the shared memory, and the kernel was slowed as a result. The supposed occupancy of a block of shared memory was 34x32 (the dimensions of the shared memory matrix) x 4 (the size of a float) which equals 4,352 bytes per block, which is supposedly less than the maximum of about 49KB stated for a device with a 5.0 compute capability (which this series of tests on individual kernel run times was performed on). With this is mind it is still unclear as to why the shared memory performed more poorly that the global memory implementation.
 +
 
 +
Unfortunately our group's inability to effectively use profiling tools has left this discrepancy as a mystery.
 +
 
 +
In conclusion, while it may be possible to parallelize the algorithm we chose well, the effort to do so would involve ensuring that shared memory is properly synchronized in two block dimensions (2 dimensions of ghost cells rather than the 1 we implemented), and to ensure that shared memory is allocated appropriately such that maximum occupancy is established within the GPU. Unfortunately, our attempts fell short, and while implementing constant memory seemed to speed up the kernel a bit, our solution was not fully scalable in both dimensions, and shared memory was not implemented in a way that improved kernel efficiency.

Latest revision as of 12:21, 10 November 2023

GPUSquad

         (           (                         (      
 (       )\ )        )\ )   (            (     )\ )   
 )\ )   (()/(    (  (()/( ( )\      (    )\   (()/(   
(()/(    /(_))   )\  /(_)))((_)     )\((((_)(  /(_))  
 /(_))_ (_))  _ ((_)(_)) ((_)_   _ ((_))\ _ )\(_))_   
(_)) __|| _ \| | | |/ __| / _ \ | | | |(_)_\(_)|   \  
  | (_ ||  _/| |_| |\__ \| (_) || |_| | / _ \  | |) | 
   \___||_|   \___/ |___/ \__\_\ \___/ /_/ \_\ |___/

Team Members

  1. TS
  2. Michael Overall
  3. Igor Krasnyanskiy
  4. Email All

Progress

Assignment 1

Idea 1 - Jacobi Method for 2D Poisson Problem

This topic is based on the following pdf: https://math.berkeley.edu/~wilken/228A.F07/chr_lecture.pdf

Background:

Poisson's equation according to Wikipedia:

"In mathematics, Poisson's equation is a partial differential equation of elliptic type with broad utility in mechanical engineering and theoretical physics. It arises, for instance, to describe the potential field caused by a given charge or mass density distribution; with the potential field known, one can then calculate gravitational or electrostatic field. It is a generalization of Laplace's equation, which is also frequently seen in physics."

According to the intro in the pdf above, finite-difference and finite-element methods are the solution techniques of choice for solving elliptic PDE problems. Regardless of which type of technique you choose, you will end up with sets of linear relationships between various variables, and will need to solve for the solution u_k which satisfies all the linear relationships prescribed by the PDE.

This can be written, according to the pdf, as a matrix "Au = b, where we wish to find a solution u, given that A is a matrix capturing the differentiation operator, and b corresponds to any source or boundary terms". The Jacobi method can be used to solve this matrix, and is used in the code sample later.

Jacobi method according to Wikipedia:

"In numerical linear algebra, the Jacobi method (or Jacobi iterative method[1]) is an algorithm for determining the solutions of a diagonally dominant system of linear equations. Each diagonal element is solved for, and an approximate value is plugged in. The process is then iterated until it converges."

**************************

CODE - Based on PDF

**************************

The following code is based on the jacobi.cc and common.cc code at the bottom of the source PDF, but there are some changes for these tests. The files use .cpp extensions instead of .cc, the code in common.cc was added to jacobi.cpp so it's a single file, and code for carrying out the Jacobi iterations in the main were placed into a seperate function called dojacobi() for gprof profiling.

The std::cout line inside the output_and_error function was commented out to avoid excessive printing to the terminal. The error_and_output function was rewritten as outputImage to get rid of terminal and file I/O during calculations, and only output the final image at the end of the calculations (this would be necessary for the parallel versions, so the changes were applied to the serial code as well to make things fair). 

Compilation on Linux:
g++ -std=c++0x -O2 -g -pg -o jacobi jacobi.cpp
// Load standard libraries
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <fstream>
#include <cmath>
#include <chrono>
using namespace std;

// Set grid size and number of iterations
const int save_iters = 20;
const int total_iters = 5000;
const int error_every = 2;
const int m = 32, n = 1024;
const float xmin = -1, xmax = 1;
const float ymin = -1, ymax = 1;

// Compute useful constants
const float pi = 3.1415926535897932384626433832795;
const float omega = 2 / (1 + sin(2 * pi / n));
const float dx = (xmax - xmin) / (m - 1);
const float dy = (ymax - ymin) / (n - 1);
const float dxxinv = 1 / (dx*dx);
const float dyyinv = 1 / (dy*dy);
const float dcent = 1 / (2 * (dxxinv + dyyinv));

// Input function
inline float f(int i, int j) {
	float x = xmin + i*dx, y = ymin + j*dy;
	return abs(x)>0.5 || abs(y)>0.5 ? 0 : 1;
}

// Common output and error routine
void output_and_error(char* filename, float *a, const int sn) {
	// Computes the error if sn%error every==0
	if (sn%error_every == 0) {
		float z, error = 0; int ij;
		for (int j = 1; j<n - 1; j++) {
			for (int i = 1; i<m - 1; i++) {
				ij = i + m*j;
				z = f(i, j) - a[ij] * (2 * dxxinv + 2 * dyyinv)
					+ dxxinv*(a[ij - 1] + a[ij + 1])
					+ dyyinv*(a[ij - m] + a[ij + m]);
				error += z*z;
			}
		}
		//cout << sn << " " << error*dx*dy << endl;
	}

	// Saves the matrix if sn<=save iters
	if (sn <= save_iters) {
		int i, j, ij = 0, ds = sizeof(float);
		float x, y, data_float; const char *pfloat;
		pfloat = (const char*)&data_float;

		ofstream outfile;
		static char fname[256];
		sprintf(fname, "%s.%d", filename, sn);
		outfile.open(fname, fstream::out
			| fstream::trunc | fstream::binary);

		data_float = m; outfile.write(pfloat, ds);

		for (i = 0; i<m; i++) {
			x = xmin + i*dx;
			data_float = x; outfile.write(pfloat, ds);
		}

		for (j = 0; j<n; j++) {
			y = ymin + j*dy;
			data_float = y;
			outfile.write(pfloat, ds);

			for (i = 0; i<m; i++) {
				data_float = a[ij++];
				outfile.write(pfloat, ds);
			}
		}

		outfile.close();
	}
}

void outputImage(char* filename, float *a) {
	// Computes the error if sn%error every==0


	// Saves the matrix if sn<=save iters
	int i, j, ij = 0, ds = sizeof(float);
	float x, y, data_float; const char *pfloat;
	pfloat = (const char*)&data_float;

	ofstream outfile;
	static char fname[256];
	sprintf(fname, "%s.%d", filename, 101);
	outfile.open(fname, fstream::out
		| fstream::trunc | fstream::binary);

	data_float = m; outfile.write(pfloat, ds);

	for (i = 0; i < m; i++) {
		x = xmin + i*dx;
		data_float = x; outfile.write(pfloat, ds);//will need to be deferred
	}

	for (j = 0; j < n; j++) {
		y = ymin + j*dy;
		data_float = y;
		outfile.write(pfloat, ds);

		for (i = 0; i < m; i++) {
			data_float = a[ij++];
			outfile.write(pfloat, ds);
		}
	}

	outfile.close();
}
void dojacobi(int i, int j, int ij, int k, float error, float u[],
	float v[], float z, float* a, float* b) {

	// Set initial guess to be identically zero
	for (ij = 0; ij<m*n; ij++) u[ij] = v[ij] = 0;
	//output_and_error("jacobi out", u, 0);

	// Carry out Jacobi iterations
	for (k = 1; k <= total_iters; k++) {
		// Alternately flip input and output matrices
		if (k % 2 == 0) { a = u; b = v; }
		else { a = v; b = u; }

		// Compute Jacobi iteration
		for (j = 1; j<n - 1; j++) {
			for (i = 1; i<m - 1; i++) {
				ij = i + m*j;
				a[ij] = (f(i, j) + dxxinv*(b[ij - 1] + b[ij + 1])
					+ dyyinv*(b[ij - m] + b[ij + m]))*dcent;
			}
		}

		// Save and compute error if necessary
		//output_and_error("jacobi out", a, k);
	}
	outputImage("jacobi out", a);

}

int main() {
	int i, j, ij, k;
	//float error, u[m*n], v[m*n], z;
	float error, z;
	float *a, *b;
	float *u;
	float *v;
	u = new float[m*n];
	v = new float[m*n];

	std::chrono::steady_clock::time_point ts, te;
	ts = std::chrono::steady_clock::now();
	
	dojacobi(i, j, ij, k, error, u, v, z, a, b);
	
	delete[] u;
	delete[] v;

	te = std::chrono::steady_clock::now();
	std::chrono::steady_clock::duration duration = te - ts;
	auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);
	std::cout << "Serial Code Time: " << ms.count() << " ms" << std::endl;

	return 0;
}
Plotting Images with gnuplot

1. cd into the folder that contains the jacobi out.# files
2. run gnuplot
3. enter: set terminal png
4. enter: set output "outfilename.png"
5. enter: plot "jacobi out.#" binary

PNG image should be created in the same folder as the jacobi out.# files

The images look something like this: m = 32, n = 1024

Dps915 gpusquad jacobi output.png

************************

Inputs / Performance

************************

In a 2D PDE such as the 2D Poisson problem, the matrix will have m * n gridpoints.

At m = 33, n = 33, iterations = 5000: 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  Ts/call  Ts/call  name    
100.00      0.05     0.05                             dojacobi(int, int, int, int, double, double*, double*, double, double*, double*)
  0.00      0.05     0.00     5001     0.00     0.00  output_and_error(char*, double*, int)
  0.00      0.05     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z16output_and_errorPcPdi

At m = 165, n = 165, iterations = 5000: 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  us/call  us/call  name    
 99.15      1.16     1.16                             dojacobi(int, int, int, int, double, double*, double*, double, double*, double*)
  0.85      1.17     0.01     5001     2.00     2.00  output_and_error(char*, double*, int)
  0.00      1.17     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z16output_and_errorPcPdi

At m = 330, n = 330, iterations = 5000: 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  us/call  us/call  name    
 99.43      5.26     5.26                             dojacobi(int, int, int, int, double, double*, double*, double, double*, double*)
  0.57      5.29     0.03     5001     6.00     6.00  output_and_error(char*, double*, int)
  0.00      5.29     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z16output_and_errorPcPdi

************************ 

The hotspot seems to be the double for-loop based on m and n in the Jacobi iterations code of the dojacobi() function. I believe these matrix calculations could be parallelized for improved performance. Note that the for-loop that the double loop is inside of is based on a constant numbers, iters, so it doesn't grow with the problem size. It would be O(iters * n^2) which is still O(n^2) not O(n^3).

Idea 2 - LZW Compression

BACKGROUND:(Paraphrased from "LZW compression" at http://whatis.techtarget.com/definition/LZW-compression) LZW compression is a compression algorithm that creates a dictionary of tokens from collections of characters. These tokens correspond to different patterns of bit values. Unique tokens are generated from the longest possible series of characters that recur in sequence any time later in a body of text. These patterns of bits that correspond to string tokens are written to an output file. Since commonly recurring sequences of words are registered to the dictionary as smaller (perhaps 12 bit) tokens, the corresponding dictionary bit code comes to represent a series of characters that would otherwise be longer than 12 bits. This is what facilitates the compression.

************************************************

CODE:

The following code is an example program that performs Compression and Decompression of text files using a LZW algorithm that encodes to 12 bit values.

This is algorithm was provided with a link as a potential project through the group projects page, but here it is again: https://codereview.stackexchange.com/questions/86543/simple-lzw-compression-algorithm

The body of code for the algorithm is as follows: 

//  Compile with gcc 4.7.2 or later, using the following command line:
//
//    g++ -std=c++0x lzw.c -o lzw
//
//LZW algorithm implemented using fixed 12 bit codes.

#include <iostream>
#include <sstream>
#include <fstream>

#include <bitset>
#include <string>
#include <unordered_map>

#define MAX_DEF 4096

using namespace std;

string convert_int_to_bin(int number)
{
    string result = bitset<12>(number).to_string();
    return result;
}

void compress(string input, int size, string filename) {
    unordered_map<string, int> compress_dictionary(MAX_DEF);
    //Dictionary initializing with ASCII
    for ( int unsigned i = 0 ; i < 256 ; i++ ){
    compress_dictionary[string(1,i)] = i;
    }
    string current_string;
    unsigned int code;
    unsigned int next_code = 256;
    //Output file for compressed data
    ofstream outputFile;
    outputFile.open(filename + ".lzw");

    for(char& c: input){
    current_string = current_string + c;
    if ( compress_dictionary.find(current_string) ==compress_dictionary.end() ){
            if (next_code <= MAX_DEF)
                compress_dictionary.insert(make_pair(current_string, next_code++));
            current_string.erase(current_string.size()-1);
            outputFile << convert_int_to_bin(compress_dictionary[current_string]);
            current_string = c;
        }   
    }   
    if (current_string.size())
            outputFile << convert_int_to_bin(compress_dictionary[current_string]);
    outputFile.close();
}



void decompress(string input, int size, string filename) {
    unordered_map<unsigned int, string> dictionary(MAX_DEF);
    //Dictionary initializing with ASCII
    for ( int unsigned i = 0 ; i < 256 ; i++ ){
    dictionary[i] = string(1,i);
    }
    string previous_string;
    unsigned int code;
    unsigned int next_code = 256;
    //Output file for decompressed data
    ofstream outputFile;
    outputFile.open(filename + "_uncompressed.txt");

    int i =0;
    while (i<size){
        //Extracting 12 bits and converting binary to decimal
        string subinput = input.substr(i,12);
        bitset<12> binary(subinput);
        code = binary.to_ullong();
        i+=12;

        if ( dictionary.find(code) ==dictionary.end() ) 
            dictionary.insert(make_pair(code,(previous_string + previous_string.substr(0,1))));
        outputFile<<dictionary[code];
        if ( previous_string.size())
            dictionary.insert(make_pair(next_code++,previous_string + dictionary[code][0])); 
        previous_string = dictionary[code];
        }
    outputFile.close();
}

string convert_char_to_string(const char *pCh, int arraySize){
    string str;
    if (pCh[arraySize-1] == '\0') str.append(pCh);
    else for(int i=0; i<arraySize; i++) str.append(1,pCh[i]);
    return str;
}

static void show_usage()
{
        cerr << "Usage: \n"
              << "Specify the file that needs to be compressed or decompressed\n"
              <<"lzw -c input    #compress file input\n"
              <<"lzw -d input    #decompress file input\n"
              <<"Compressed data will be found in a file with the same name but with a .lzw extension\n"
              <<"Decompressed data can be found in a file with the same name and a _uncompressed.txt extension\n"
              << endl;
}


int main (int argc, char* argv[]) {
    streampos size;
    char * memblock;

    if (argc <2)
    {
        show_usage();   
        return(1);
    }
    ifstream file (argv[2], ios::in|ios::binary|ios::ate);
    if (file.is_open())
    {
        size = file.tellg();
        memblock = new char[size];
        file.seekg (0, ios::beg);
        file.read (memblock, size);
        file.close();
        string input = convert_char_to_string(memblock,size);
        if (string( "-c" ) == argv[1] )
            compress(input,size, argv[2]);
        else if (string( "-d" ) == argv[1] )
            decompress(input,size, argv[2]);
        else
            show_usage();
    }
    else {
    cout << "Unable to open file."<<endl;
    show_usage();
    }
    return 0;
}


PROFILING:

The above program needs an input file to compress and decompress text in. For the purposes of testing, the Gutenberg press' "Complete Works of Shakespeare" was used as an input text file (http://www.gutenberg.org/files/100/100-0.txt) because it represents a large enough body of text to actually have perceptible run times for compression. Increases in the size of the data used are created through copying one full version of the text in the last iteration of testing and appending it to the end of the text file (so one Shakespeare's complete works becomes two back to back, two becomes four, etc).

PROFILING WITH THE ORIGINAL TEXT:

[[File:Flat profile:

Each sample counts as 0.01 seconds. 

 %   cumulative   self              self     total           
time   seconds   seconds    calls  ns/call  ns/call  name    
46.15      0.24     0.24                             compress(std::string, int, std::string)
25.00      0.37     0.13  7954538    16.34    16.34  show_usage()
21.15      0.48     0.11  2091647    52.59    52.59  convert_int_to_bin(int)
 7.69      0.52     0.04  2091647    19.12    35.47  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string const&)
 0.00      0.52     0.00     3841     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, unsigned int> >(std::pair<std::string, unsigned int>&&, unsigned int, unsigned int)
 0.00      0.52     0.00      256     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, int> >(std::pair<std::string, int>&&, unsigned int, unsigned int)
 0.00      0.52     0.00      256     0.00    16.34  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string&&)
 0.00      0.52     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z18convert_int_to_bini]]


PROFILING WITH TWICE THE TEXT:

Flat profile:

Each sample counts as 0.01 seconds. 

 %   cumulative   self              self     total           
time   seconds   seconds    calls  ns/call  ns/call  name    
38.71      0.24     0.24                             compress(std::string, int, std::string)
25.81      0.40     0.16  2091647    76.49    76.49  convert_int_to_bin(int)
22.58      0.54     0.14  7954538    17.60    17.60  show_usage()
 9.68      0.60     0.06  2091647    28.69    46.29  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string const&)
 3.23      0.62     0.02                             convert_char_to_string(char const*, int)
 0.00      0.62     0.00     3841     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, unsigned int> >(std::pair<std::string, unsigned int>&&, unsigned int, unsigned int)
 0.00      0.62     0.00      256     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, int> >(std::pair<std::string, int>&&, unsigned int, unsigned int)
 0.00      0.62     0.00      256     0.00    17.60  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string&&)
 0.00      0.62     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z18convert_int_to_bini


PROFILING WITH 4X THE TEXT:

Flat profile:

Each sample counts as 0.01 seconds. 

 %   cumulative   self              self     total           
time   seconds   seconds    calls  us/call  us/call  name    
40.48      1.02     1.02                             compress(std::string, int, std::string)
30.16      1.78     0.76 31802927     0.02     0.02  show_usage()
22.62      2.35     0.57  8363660     0.07     0.07  convert_int_to_bin(int)
 5.56      2.49     0.14  8363660     0.02     0.04  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string const&)
 0.79      2.51     0.02                             convert_char_to_string(char const*, int)
 0.40      2.52     0.01      256    39.06    39.09  std::__detail::_Map_base<std::string, std::pair<std::string const, int>, std::_Select1st<std::pair<std::string const, int> >, true, std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true> >::operator[](std::string&&)
 0.00      2.52     0.00     3841     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, unsigned int> >(std::pair<std::string, unsigned int>&&, unsigned int, unsigned int)
 0.00      2.52     0.00      256     0.00     0.00  std::__detail::_Hashtable_iterator<std::pair<std::string const, int>, false, false> std::_Hashtable<std::string, std::pair<std::string const, int>, std::allocator<std::pair<std::string const, int> >, std::_Select1st<std::pair<std::string const, int> >, std::equal_to<std::string>, std::hash<std::string>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, false, false, true>::_M_insert_bucket<std::pair<std::string, int> >(std::pair<std::string, int>&&, unsigned int, unsigned int)
 0.00      2.52     0.00        1     0.00     0.00  _GLOBAL__sub_I__Z18convert_int_to_bini


A useful hotspot for parallelization is not immediately obvious through profiling, since the main compress() function contains the bulk of the program logic, and its show_usage(, and convert_int_to_bin are simple functions, that are called frequently in one for loop.

What really affects the runtime of the program based on data is the extent to which long matching strings can be tokenized. When this happens, larger, and larger chunks of text can be processed and compressed in an iteration. This is reflected by the fact that pasting the same blocks of text over again to increase data size does not proportionally increase run time because the same tokens work for subsequent pasted blocks.


POTENTIAL FOR PARALLELIZATION:

The compress() function performs similar operations on a collection of text, however it relies on a dictionary and an expanding string to be tokenized in a dictrionary. This could potentially be paralellized through a divide and conquer strategy where gpu blocks with shared caches share their own dictionaries and iterate over their own block of text. This, however would not be particularly useful because LZW compression relies on a globally accessible dictionary that can be continuously modified. Having multiple threads try to access the same dictionary tokens in global memory would create race conditions, and creating block-local dictionaries in shared memory would reduce the efficacy of having a large globally available dictionary to tokenize strings.

What all of this means is that LZW compression would probably be a poor candidate for parallelization due to the way that parallel memory would try to update the token dictionary.

Idea 3 - MergeSort

Based on assignment topic suggestions.

Code for logic is based on BTP500 course note by Catherine Leung - https://cathyatseneca.gitbooks.io/data-structures-and-algorithms/content/sorting/merge_sort_code.html

It has been adjust to work only with int arrays and perform operation on worst case scenario for merge sort, which is something similar to this array - [0, 2, 4, 6, 1, 3, 5, 7].


CODE

To compile on matrix - g++ -O2 -g -pg -oa1 a1.cpp

It performs merge sort on array with 100000000 which has worst case scenario positioning of elements for sorting. 

#include<iostream>

int SIZE = 100000000;
/*This function merges the two halves of the array arr into tmp and then copies it back into arr*/
void merge(int arr[], int tmp[], int startA, int startB, int end) {
	int aptr = startA;
	int bptr = startB;
	int i = startA;
	while (aptr<startB && bptr <= end) {
		if (arr[aptr]<arr[bptr])
			tmp[i++] = arr[aptr++];
		else
			tmp[i++] = arr[bptr++];
	}
	while (aptr<startB) {
		tmp[i++] = arr[aptr++];
	}
	while (bptr <= end) {
		tmp[i++] = arr[bptr++];
	}
	for (i = startA; i <= end; i++) {
		arr[i] = tmp[i];
	}
}

//this function sorts arr from index start to end
void mSort(int* arr, int* tmp, int start, int end) {
	if (start<end) {
		int mid = (start + end) / 2;
		mSort(arr, tmp, start, mid);
		mSort(arr, tmp, mid + 1, end);
		merge(arr, tmp, start, mid + 1, end);
	}
}

void mergeSort(int* arr, int size) {
	int* tmp = new int[size];
	mSort(arr, tmp, 0, size - 1);
	delete[] tmp;
}

void printArr(int* arr) {
	for (int i = 0; i < SIZE; i++) {
		std::cout << arr[i] << std::endl;
	}
}
int main(){
	int* sampleArr = new int[SIZE];
	bool worstCaseFlag = false;
	int j = 1;
	for (int i = 0; i < SIZE; i++) {
		if (worstCaseFlag == false) {
			sampleArr[i] = i;
			worstCaseFlag = true;
		}
		else {
			sampleArr[i] = j;
			j += 2;
			worstCaseFlag = false;
		}
	}
	mergeSort(sampleArr, SIZE);
	//printArr(sampleArr);
	std::cin.get();
	return 0;
}


Profiling

Profiling with 1000000 elements 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls  ms/call  ms/call  name
 69.23      0.09     0.09   999999     0.00     0.00  merge(int*, int*, int, int, int)
 30.77      0.13     0.04        1    40.00   130.00  mSort(int*, int*, int, int)
  0.00      0.13     0.00        1     0.00     0.00  _GLOBAL__sub_I_SIZE


Profiling with 10000000 elements 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 88.89      1.04     1.04  9999999     0.00     0.00  merge(int*, int*, int, int, int)
 11.11      1.17     0.13        1     0.13     1.17  mSort(int*, int*, int, int)
  0.00      1.17     0.00        1     0.00     0.00  _GLOBAL__sub_I_SIZE


Profiling with 100000000 elements 

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 90.40     13.85    13.85 99999999     0.00     0.00  merge(int*, int*, int, int, int)
  9.60     15.32     1.47        1     1.47    15.32  mSort(int*, int*, int, int)
  0.00     15.32     0.00        1     0.00     0.00  _GLOBAL__sub_I_SIZE

Analysing


The most time consuming part is merging, which can be at least partially paralleled. The Big(O) of this case is O(n).

Assignment 2

We chose to move forward with the Jacobi Method for 2D Poisson Equations code.

We parallelized the original code by placing the jacobi calculations into a kernel. For this initial parallel version, we only used 1D threading and had each thread run a for loop for the other dimension.

The iters loop launches a kernel for each iteration and we use double buffering (where we choose to launch the kernel with either d_a, d_b or d_b, d_a) since we can't simply swap pointers like in the serial code.

Double buffering is needed so that there is a static version of the the matrix before the current set of calculations are done. Since an element of the matrix is calculated based on elements on all four sides, if one of those elements updated itself while calculations were being done on another element, you would end up with the wrong answer/race conditions.

****************

CODE

**************** 

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include <stdio.h>

// Load standard libraries
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <fstream>
#include <cmath>
#include <chrono>
using namespace std;
__global__ void matrixKernel(float* a, float* b, float dcent, int n, int m, int xmin, int xmax,
	int ymin, int ymax, float dx, float dy, float dxxinv, float dyyinv) {//MODIFY to suit algorithm

	int j = blockDim.x * blockIdx.x + threadIdx.x + 1;

	//above: we are using block and thread indexes to replace some of the iteration logic
	if (j < n - 1) {
		for (int i = 1; i < m - 1; i++) {
			int  ij = i + m*j;

			float x = xmin + i*dx, y = ymin + j*dy;

			float input = abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;

			a[ij] = (input + dxxinv*(b[ij - 1] + b[ij + 1])
				+ dyyinv*(b[ij - m] + b[ij + m]))*dcent;
		}
	}

}

// Set grid size and number of iterations
const int save_iters = 20;
const int total_iters = 5000;
const int error_every = 2;
const int m = 32, n = 1024;
const float xmin = -1, xmax = 1;
const float ymin = -1, ymax = 1;

// Compute useful constants
const float pi = 3.1415926535897932384626433832795;
const float omega = 2 / (1 + sin(2 * pi / n));
const float dx = (xmax - xmin) / (m - 1);
const float dy = (ymax - ymin) / (n - 1);
const float dxxinv = 1 / (dx*dx);
const float dyyinv = 1 / (dy*dy);
const float dcent = 1 / (2 * (dxxinv + dyyinv));

// Input function
inline float f(int i, int j) {
	float x = xmin + i*dx, y = ymin + j*dy;
	return abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
}

// Common output and error routine
void outputImage(char* filename, float *a) {
	// Computes the error if sn%error every==0


	// Saves the matrix if sn<=save iters
	int i, j, ij = 0, ds = sizeof(float);
	float x, y, data_float; const char *pfloat;
	pfloat = (const char*)&data_float;

	ofstream outfile;
	static char fname[256];
	sprintf(fname, "%s.%d", filename, 101);
	outfile.open(fname, fstream::out
		| fstream::trunc | fstream::binary);

	data_float = m; outfile.write(pfloat, ds);

	for (i = 0; i < m; i++) {
		x = xmin + i*dx;
		data_float = x; outfile.write(pfloat, ds);
	}

	for (j = 0; j < n; j++) {
		y = ymin + j*dy;
		data_float = y;
		outfile.write(pfloat, ds);

		for (i = 0; i < m; i++) {
			data_float = a[ij++];
			outfile.write(pfloat, ds);
		}
	}

	outfile.close();
}


void dojacobi() {
	int i, j, ij, k;
	float error, z;
	float *a, *b;
	float *u;
	float *v;
	u = new float[m*n];
	v = new float[m*n];

	// Set initial guess to be identically zero
	for (ij = 0; ij < m*n; ij++) u[ij] = v[ij] = 0;

	a = v; b = u;
	float* d_a;
	float* d_b;

	//malloc
	cudaMalloc((void**)&d_a, m*n * sizeof(float));
	cudaMalloc((void**)&d_b, m*n * sizeof(float));

	cudaMemcpy(d_a, a, n* m * sizeof(float), cudaMemcpyHostToDevice);
	cudaMemcpy(d_b, b, n* m * sizeof(float), cudaMemcpyHostToDevice);

	int nblocks = n / 1024;
	dim3 dGrid(nblocks);
	dim3 dBlock(1024);
	// Carry out Jacobi iterations
	for (k = 1; k <= total_iters; k++) {
		if (k % 2 == 0) {
			cudaError_t error = cudaGetLastError();

			matrixKernel << <dGrid, dBlock >> > (d_a, d_b, dcent, n, m, xmin, xmax, ymin, ymax, dx, dy, dxxinv, dyyinv);

			cudaDeviceSynchronize();
			if (cudaGetLastError()) {
				std::cout << "error";
			}
		}

		else {
			cudaError_t error = cudaGetLastError();

			matrixKernel << <dGrid, dBlock >> > (d_b, d_a, dcent, n, m, xmin, xmax, ymin, ymax, dx, dy, dxxinv, dyyinv);

			cudaDeviceSynchronize();
			if (cudaGetLastError()) {
				std::cout << "error";
			}
		}
	}

	cudaMemcpy(a, d_a, n* m * sizeof(float), cudaMemcpyDeviceToHost);
	outputImage("jacobi out", a);
	cudaFree(d_a);
	cudaFree(d_b);
	delete[] u;
	delete[] v;
}

int main() {
	std::chrono::steady_clock::time_point ts, te;
	ts = std::chrono::steady_clock::now();
	dojacobi();
	te = std::chrono::steady_clock::now();
	std::chrono::steady_clock::duration duration = te - ts;
	auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);
	std::cout << "Parallel Code Time: " << ms.count() << " ms" << std::endl;
	cudaDeviceReset();

	return 0;
}

****************

A2 TIMINGS

****************

Dps915 gpusquad a2 chart.png

Assignment 3

Optimization techniques attempted

  • Get rid of the for loop in the kernel and use 2D threading within blocks
  • Use gpu constant memory for jacobi calculation constants
  • Utilize the ghost cell pattern for shared memory within blocks

The ghost cell pattern is a technique used to allow threads in a particular block to access data that would normally be inside another block. We make a shared memory array larger than the number of threads per block in a particular dimension, then when we are at the first or last thread, can copy a neighbouring cells' data based on a global index, to our local ghost cell.

The following info and images are taken from: http://people.csail.mit.edu/fred/ghost_cell.pdf

Ghost Cell Pattern Abstract: 

Many problems consist of a structured grid of points that
are updated repeatedly based on the values of a fixed set
of neighboring points in the same grid. To parallelize these
problems we can geometrically divide the grid into chunks
that are processed by different processors. One challenge
with this approach is that the update of points at the periphery
of a chunk requires values from neighboring chunks.
These are often located in remote memory belonging to different
processes. The naive implementation results in a lot
of time spent on communication leaving less time for useful
computation. By using the Ghost Cell Pattern communication
overhead can be reduced. This results in faster time to
completion.

In our code, our main jacobi calculations look like this (when using global indexing): 

a[ij] = (input + dxxinv*(b[ij - 1] + b[ij + 1])
	+ dyyinv*(b[ij - m] + b[ij + m]))*dcent;
  • b[ij - 1] is a cell's left neighbour
  • b[ij + 1] is a cell's right neighbour
  • b[ij - m] is a cell's top neighbour (we represent a 2D array as a 1D array, so subtract row length to go one cell up)
  • b[ij + m] is a cell's bottom neighbour (add row to go one cell down)

Dps 915 gpusquad needneighbour.png

Dps 915 gpusquad ghostcellexample.png

We use a 1D grid of blocks which have 2D threads. Thus the first block needs ghost cells on the right side, the last block needs ghost cells on the left side, and all other blocks in the middle need ghost cells on both the left and right sides.

In A2 and A3 we only scaled the columns (we would have, had we not mixed up the indexing--the 2d version scales along the n dimension, while the 1d versions scale properly along the n dimension), and the other dimension always stayed at 32. This made it easy for us to keep the grid of blocks one dimensional. If we scaled the rows as well as the columns we would most likely have need to use a two dimensional grid of blocks. This would make the problem more difficult as we would then not only need ghost columns, but also ghost rows on the top and bottom of blocks. This would likely cause synchronization problems because we would need to make sure that block execution was synchronized along 2 dimensions instead of just one.

Dps915 gpusquad 2Dghostcells.png

****************

CODE FOR 1D GRID WITH 2D THREAD SETUP (kernel contains shared and global memory versions)

**************** 

#include <cuda.h>
#include "cuda_runtime.h"
#include <device_launch_parameters.h>
#include <stdio.h>
#include <device_functions.h>
#include <cuda_runtime_api.h>
#ifndef __CUDACC__
#define __CUDACC__
#endif
// Load standard libraries
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <fstream>
#include <cmath>
#include <chrono>
using namespace std;
const int ntpb = 1024;//TOTAL number of threads per block
const int ntpbXY = 32;
__device__ __constant__ int d_m;
__device__ __constant__ int d_n;
__device__ __constant__ float d_xmin;
__device__ __constant__ float d_xmax;
__device__ __constant__ float d_ymin;
__device__ __constant__ float d_ymax;
__device__ __constant__ float d_pi;
__device__ __constant__ float d_omega;
__device__ __constant__ float d_dx;
__device__ __constant__ float d_dy;
__device__ __constant__ float d_dxxinv;
__device__ __constant__ float d_dyyinv;
__device__ __constant__ float d_dcent;
__device__ __constant__ int d_nBlocks;
__device__ __constant__ int d_ntpbXY;
void check(cudaError_t err) {
  if (err != cudaSuccess)
    std::cerr << "ERROR: *** " << cudaGetErrorString(err) << " ****" << std::endl;
}
//every thread does a single calculation--remove iterations
//-take everything from one dimensional global, and put into 2d shared--shared has ghost b/c need for threads--
//--global is 1d indexed ,shared is 2d indexed. 
//TODO:  remove for loop from kernel and index entirely with blockIdx, blockDim, and threadIdx (check workshop 7 for indexing by shared memory within blocks)
__global__ void matrixKernel(float* a, float* b) {//MODIFY to suit algorithm
    //int j= blockDim.x*blockIdx.x + threadIdx.x+1;//TODO: CHECK THIS DOESN'T MESS UP GLOBAL EXECUTION
  int j = threadIdx.y;
  int i = blockDim.x*blockIdx.x + threadIdx.x;
    //int i = threadIdx.y + 1;//original code has k = 1 and <=total_iters--since k will now start at 0, set end at < total_iters
                                                      //above: we are using block and thread indexes to replace some of the iteration logic
  int tj = threadIdx.y;
  int ti = threadIdx.x;
  int tij = threadIdx.x + ntpbXY*threadIdx.y;

    
  int  ij = i + d_m*j;

  //GLOBAL MEMORY VERSION (COMMENT OUT SHARED SECTION BELOW TO RUN):==================================
//  if (!(blockIdx.x == 0 && threadIdx.x == 0 || blockIdx.x == (d_nBlocks - 1) && threadIdx.x == (d_ntpbXY - 1)) && threadIdx.y != 0 && threadIdx.y != (d_ntpbXY - 1)) {
 //   float x = d_xmin + i*d_dx, y = d_ymin + j*d_dy;
  //  float input = abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
   // a[ij] = (input + d_dxxinv*(b[ij - 1] + b[ij + 1])
    //  + d_dyyinv*(b[ij - d_m] + b[ij + d_m]))*d_dcent;
//  }


//SHARED MEMORY SETUP (COMMENT OUT FOR GLOBAL VERSION)
  __shared__ float bShared[ntpbXY+2][ntpbXY];
  //NOTES ON SHARED INDEX OFFSET:
  //x offset is -1, so to get the current index for the thread for global, it is bShared[ti+i][tj]
  //-to get global x index -1 is bShared[ti][tj]
  //-to get global x index is bShared[ti+1][tj]
  //-to get global x index +1 is bShared[ti+2][tj]

  //set left ghost cells
  if (threadIdx.x == 0 && blockIdx.x != 0)
    bShared[ti][tj] = b[ij - 1];

  //set right ghost cells
  if (threadIdx.x == ntpbXY - 1 && blockIdx.x != d_nBlocks - 1)
    bShared[ti+2][tj] = b[ij+1];

  //set ghost cell for current thread relative to global memory
  bShared[ti+1][tj] = b[d_m* j + i];

  __syncthreads();

 
  //SHARED MEMORY VERSION:==========
  if (!(blockIdx.x == 0 && threadIdx.x == 0 || blockIdx.x == (d_nBlocks - 1) && threadIdx.x == (d_ntpbXY - 1)) && threadIdx.y != 0 && threadIdx.y != (d_ntpbXY - 1)) {
    float x = d_xmin + i*d_dx, y = d_ymin + j*d_dy;
    float input = abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
        a[d_m* j + i] = (input + d_dxxinv*(bShared[ti][tj] + bShared[ti+2][tj])//TODO: does the program logic allow for this to go out of range??
    + d_dyyinv*(bShared[ti+1][tj-1] + bShared[ti+1][tj+1]))*d_dcent;
      __syncthreads();
     
    }
}

// Set grid size and number of iterations
const int save_iters = 20;
const int total_iters = 5000;
const int error_every = 2;
//HERE:
const int m = 32, n =32 ;
const float xmin = -1, xmax = 1;
const float ymin = -1, ymax = 1;
// Compute useful constants
const float pi = 3.1415926535897932384626433832795;
const float omega = 2 / (1 + sin(2 * pi / n));
const float dx = (xmax - xmin) / (m - 1);//TODO: modify to fit kernel?
const float dy = (ymax - ymin) / (n - 1);//TODO: modify to fit kernel?
const float dxxinv = 1 / (dx*dx);
const float dyyinv = 1 / (dy*dy);
const float dcent = 1 / (2 * (dxxinv + dyyinv));
// Input function
inline float f(int i, int j) {
    float x = xmin + i*dx, y = ymin + j*dy;
    return abs(x) > 0.5 || abs(y) > 0.5 ? 0 : 1;
}
// Common output and error routine
void outputImage(char* filename, float *a) {
    // Computes the error if sn%error every==0
    // Saves the matrix if sn<=save iters
    int i, j, ij = 0, ds = sizeof(float);
    float x, y, data_float; const char *pfloat;
    pfloat = (const char*)&data_float;
    ofstream outfile;
    static char fname[256];
    sprintf(fname, "%s.%d", filename, 101);
    outfile.open(fname, fstream::out
        | fstream::trunc | fstream::binary);
    data_float = m; outfile.write(pfloat, ds);
    for (i = 0; i < m; i++) {
        x = xmin + i*dx;
        data_float = x; outfile.write(pfloat, ds);
    }
    for (j = 0; j < n; j++) {
        y = ymin + j*dy;
        data_float = y;
        outfile.write(pfloat, ds);
        for (i = 0; i < m; i++) {
            data_float = a[ij++];
            outfile.write(pfloat, ds);
        }
    }
    outfile.close();
}
void dojacobi() {
    int i, j, ij, k;
    float error, z;
    float *a, *b;
    float *u;
    float *v;
    u = new float[m*n];
    v = new float[m*n];
    // Set initial guess to be identically zero
    for (ij = 0; ij < m*n; ij++) u[ij] = v[ij] = 0;
    a = v; b = u;
    float* d_a;
    float* d_b;
    int nBlocks = ((n*m + ntpb - 1) / ntpb);
    //malloc
    cudaMalloc((void**)&d_a, m*n * sizeof(float));
    cudaMalloc((void**)&d_b, m*n * sizeof(float));
    cudaMemcpy(d_a, a, n* m * sizeof(float), cudaMemcpyHostToDevice);
    cudaMemcpy(d_b, b, n* m * sizeof(float), cudaMemcpyHostToDevice);
    dim3 dGrid(nBlocks, 1);//1D grid of blocks (means we don't have to ghost 1st and last rows of matrices w/i a block
    dim3 dBlock(32, 32);
  cudaMemcpyToSymbol(d_nBlocks, &nBlocks, sizeof(int));
  cudaMemcpyToSymbol(d_ntpbXY, &ntpbXY, sizeof(int));
    // Carry out Jacobi iterations
    for (k = 1; k <= total_iters; k++) {
        if (k % 2 == 0) {
            cudaError_t error = cudaGetLastError();
//RUN KERNEL:
            matrixKernel << <dGrid, dBlock >> > (d_a, d_b);
            cudaDeviceSynchronize();
      check(cudaGetLastError());
        }
        else {
            cudaError_t error = cudaGetLastError();
            matrixKernel << <dGrid, dBlock >> > (d_b, d_a);
            cudaDeviceSynchronize();
            if (cudaGetLastError()) {
                std::cout << "error";
            }
        }
    }
    cudaMemcpy(a, d_a, n* m * sizeof(float), cudaMemcpyDeviceToHost);
    outputImage("jacobi out", a);
    cudaFree(d_a);
    cudaFree(d_b);
    delete[] u;
    delete[] v;
}
int main() {
    cudaError_t cuerr;
    
    cuerr = cudaMemcpyToSymbol(d_m, &m, sizeof(int));
    if (cudaGetLastError()) {
        std::cout << cudaGetErrorString(cuerr) << std::endl;
    }
    cudaMemcpyToSymbol(d_n, &n, sizeof(int));
    cudaMemcpyToSymbol(d_xmin, &xmin, sizeof(float));
    cudaMemcpyToSymbol(d_xmax, &xmax, sizeof(float));
    cudaMemcpyToSymbol(d_ymin, &ymin, sizeof(float));
    cudaMemcpyToSymbol(d_ymax, &ymax, sizeof(float));
    cudaMemcpyToSymbol(d_pi, &pi, sizeof(float));
    cudaMemcpyToSymbol(d_omega, &pi, sizeof(float));
    cudaMemcpyToSymbol(d_dx, &dx, sizeof(float));
    cudaMemcpyToSymbol(d_dy, &dy, sizeof(float));

    cuerr = cudaMemcpyToSymbol(d_dxxinv, &dxxinv, sizeof(float));
    if (cudaGetLastError()) {
        std::cout << "error" << std::endl;
        std::cout << cudaGetErrorString(cuerr) << std::endl;
    }
    cudaMemcpyToSymbol(d_dyyinv, &dyyinv, sizeof(float));
    cuerr = cudaMemcpyToSymbol(d_dcent, &dcent, sizeof(float));
    if (cudaGetLastError()) {
        //std::cout << "error";
        std::cout << cudaGetErrorString(cuerr) << std::endl;
    }
    std::chrono::steady_clock::time_point ts, te;
    ts = std::chrono::steady_clock::now();
    dojacobi();
    te = std::chrono::steady_clock::now();
    std::chrono::steady_clock::duration duration = te - ts;
    auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);
    std::cout << "Parallel Code Time: " << ms.count() << " ms" << std::endl;
    cudaDeviceReset();
    return 0;
}

****************

A NOTE ON SCALABILITY:

In our attempts to make the kernel scalable with ghost cells, we scaled along one dimension. However, we were inconsistent in our scaling. The 1D kernel scaled along the n (y) dimension while the 2d kernels scaled along the m (x) dimension. Scaling along the x dimension, while allowing results to be testable between serial and 2D parallelized versions of the code, produced distributions that were strangely banded and skewed. In other words, we made the code render weird things faster:

MDimensionScale.png

FINAL TIMINGS 
 THE GRAPH IMMEDIATELY BELOW IS INCORRECT: there was an error recording the 1D runtimes for assignment 2

****************

Dps915 gpusquad a3chart.png

****************

PROPER TIMINGS:

Code timings2.png

Blue = Serial (A1) Orange = Parallel (A2) Grey = Optimized Global (A3) Yellow = Optimized Shared (A3)

The above graph includes the total run times for the serial code, the 1D kernel from assignment 2, a kernel with global and constant memory with a 2D thread arrangement, and the same 2D arrangement but with shared memory utilizing ghost cells.

We found that the most efficient version of the code was the 2d version that used constant memory and did not use shared memory. Because the shared memory version of the kernel required synchronization of threads to allocate shared memory every time a kernel was run, and a kernel was run 5000 times for each version of our code, this increased overhead for memory setup actually made the execution slower than the version with global memory.

We found that the most efficient version of the code was the 2d version that used constant memory and did not use shared memory. Because the shared memory version of the kernel required synchronization of threads to allocate shared memory every time a kernel was run, and a kernel was run 5000 times for each version of our code, the if statements required to set up the ghost cells for shared memory may have created a certain amount of warp divergence, thus slowing down the runtimes of each individual kernel.

Below, are two images that show 4 consecutive kernel runs for both global and shared versions of the code. It is apparent that shared kernel runs actually take more time than the global memory versions.


TIMES FOR THE GLOBAL KERNEL KernelGlobalTimes.png


TIMES FOR THE SHARED KERNEL

SharedKernelTimes.png

Note how the run times for each kernel with shared memory are significantly longer than those with global.

To try to determine if this issue was one of warp divergence, we tried to time a kernel with global memory that also initialized shared memory, although referenced global memory when carrying out the actual calculations:

GlobalInitSharedKernelTimes.png

The run of a kernel that allocated shared memory using a series of if statements, but executed instructions using global memory is shown in the figure above. While slightly longer than the run with global memory where shared memory is not initialized for ghost cells, it still takes less time to run than the version with Global memory. It is likely that Our group's attempts to employ shared memory failed because we did not adequately schedule or partition the shared memory, and the kernel was slowed as a result. The supposed occupancy of a block of shared memory was 34x32 (the dimensions of the shared memory matrix) x 4 (the size of a float) which equals 4,352 bytes per block, which is supposedly less than the maximum of about 49KB stated for a device with a 5.0 compute capability (which this series of tests on individual kernel run times was performed on). With this is mind it is still unclear as to why the shared memory performed more poorly that the global memory implementation.

Unfortunately our group's inability to effectively use profiling tools has left this discrepancy as a mystery.

In conclusion, while it may be possible to parallelize the algorithm we chose well, the effort to do so would involve ensuring that shared memory is properly synchronized in two block dimensions (2 dimensions of ghost cells rather than the 1 we implemented), and to ensure that shared memory is allocated appropriately such that maximum occupancy is established within the GPU. Unfortunately, our attempts fell short, and while implementing constant memory seemed to speed up the kernel a bit, our solution was not fully scalable in both dimensions, and shared memory was not implemented in a way that improved kernel efficiency.

Retrieved from "https://wiki.cdot.senecapolytechnic.ca/w/index.php?title=GPUSquad&oldid=160854"