You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
279 lines
8.7 KiB
279 lines
8.7 KiB
#include <includes.hpp>
|
|
#include <iomanip>
|
|
#include <utilities.hpp>
|
|
#include <boost/program_options.hpp>
|
|
#include <bsp.hpp>
|
|
|
|
#include "wavelet.hpp"
|
|
#include "wavelet_parallel.hpp"
|
|
|
|
static unsigned int P;
|
|
static unsigned int W;
|
|
static unsigned int H;
|
|
|
|
static std::vector<double> data;
|
|
static std::vector<double> seqr;
|
|
static std::vector<double> parr;
|
|
|
|
using namespace wvlt;
|
|
|
|
struct plan_2D {
|
|
par::plan_1D horizontal;
|
|
par::plan_1D vertical;
|
|
};
|
|
|
|
struct block {
|
|
std::vector<double> data;
|
|
std::vector<double> hcomm;
|
|
std::vector<double> vcomm;
|
|
|
|
block(plan_2D const & plan)
|
|
: data(plan.horizontal.b * plan.vertical.b, 0.0)
|
|
, hcomm(plan.horizontal.Cm * plan.vertical.b, 0.0)
|
|
, vcomm(plan.horizontal.b * plan.vertical.Cm, 0.0)
|
|
{}
|
|
|
|
void push() {
|
|
bsp::push_reg(data.data(), data.size());
|
|
bsp::push_reg(hcomm.data(), hcomm.size());
|
|
bsp::push_reg(vcomm.data(), vcomm.size());
|
|
}
|
|
};
|
|
|
|
// Communicate vertical data from b (strided) to b2 (not strided)
|
|
static void vcomm_step(wvlt::par::proc_info const & pi, plan_2D const & plan, block const & b, block & b2, unsigned int stride){
|
|
const unsigned int Cm = plan.vertical.Cm;
|
|
const unsigned int width = plan.horizontal.b;
|
|
for(unsigned int i = 0; i < Cm; ++i){
|
|
for(unsigned int x = 0; x < width; ++x){
|
|
bsp::put(pi.prev, b.data.data() + x + width*stride*i, b2.vcomm.data(), x + width*i, 1);
|
|
}
|
|
}
|
|
}
|
|
|
|
// Communicate horizontal data from b (strided) to b2 (not strided)
|
|
static void hcomm_step(wvlt::par::proc_info const & pi, plan_2D const & plan, block const & b, block & b2, unsigned int stride){
|
|
const unsigned int Cm = plan.horizontal.Cm;
|
|
const unsigned int width = plan.horizontal.b;
|
|
for(unsigned int y = 0; y < plan.vertical.b; ++y){
|
|
for(unsigned int i = 0; i < Cm; ++i){
|
|
bsp::put(pi.prev, b.data.data() + width*y + stride*i, b2.hcomm.data(), Cm*y + i, 1);
|
|
}
|
|
}
|
|
}
|
|
|
|
static void vcomp_step(plan_2D const & plan, block & b, unsigned int stride){
|
|
const unsigned int width = plan.horizontal.b;
|
|
const unsigned int height = plan.vertical.b;
|
|
unsigned int end = pow_two(plan.vertical.m);
|
|
for(unsigned int x = 0; x < width; ++x){
|
|
for(unsigned int i = 1; i < end; i <<= 1){
|
|
wavelet_mul(b.data.data() + x, b.vcomm[x], b.vcomm[x + width*i], width*height, width*stride*i);
|
|
if(i < end/2) wavelet_mul_base(&b.vcomm[x], width*(2*end - 2*i), width*i);
|
|
}
|
|
}
|
|
}
|
|
|
|
static void hcomp_step(plan_2D const & plan, block & b, unsigned int stride){
|
|
const unsigned int width = plan.horizontal.b;
|
|
const unsigned int height = plan.vertical.b;
|
|
unsigned int end = pow_two(plan.horizontal.m);
|
|
for(unsigned int y = 0; y < height; ++y){
|
|
for(unsigned int i = 1; i < end; i <<= 1){
|
|
auto x0 = b.hcomm[plan.horizontal.Cm*y];
|
|
auto x1 = b.hcomm[plan.horizontal.Cm*y + i];
|
|
wavelet_mul(b.data.data() + width*y, x0, x1, width, stride*i);
|
|
if(i < end/2) wavelet_mul_base(&b.hcomm[plan.horizontal.Cm*y], 2*end - 2*i, i);
|
|
}
|
|
}
|
|
}
|
|
|
|
static void vstep(wvlt::par::proc_info const & pi, plan_2D const & plan, std::vector<block> & blocks, unsigned int stride){
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int b2 = b;
|
|
vcomm_step(pi, plan, blocks[b], blocks[b2], stride);
|
|
}
|
|
bsp::sync();
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
vcomp_step(plan, blocks[b], stride);
|
|
}
|
|
}
|
|
|
|
static void hstep(wvlt::par::proc_info const & pi, plan_2D const & plan, std::vector<block> & blocks, unsigned int stride){
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int b2 = (b-1+blocks.size()) % blocks.size();
|
|
hcomm_step(pi, plan, blocks[b], blocks[b2], stride);
|
|
}
|
|
bsp::sync();
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
hcomp_step(plan, blocks[b], stride);
|
|
}
|
|
}
|
|
|
|
static void par_wavelet_2D(){
|
|
bsp::begin(P);
|
|
|
|
const wvlt::par::proc_info d(bsp::nprocs(), bsp::pid());
|
|
const wvlt::par::plan_1D horizontal(W, W/d.p, 1);
|
|
const wvlt::par::plan_1D vertical(H, H/d.p, 1);
|
|
const plan_2D plan{horizontal, vertical};
|
|
|
|
auto bbb = block(plan);
|
|
std::vector<block> blocks(d.p, bbb);
|
|
std::vector<double> hfinish(2 * d.p * vertical.b, 0.0);
|
|
std::vector<double> vfinish(horizontal.b * 2 * d.p, 0.0);
|
|
|
|
// Direct read because MCBSP can do this ;D
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int x_start = b * horizontal.b;
|
|
unsigned int y_start = (d.s - b + d.p)%d.p * vertical.b;
|
|
for(unsigned int y = 0; y < vertical.b; ++y){
|
|
for(unsigned int x = 0; x < horizontal.b; ++x){
|
|
auto v = data[x_start + x + horizontal.n*(y_start+y)];
|
|
blocks[b].data[x + horizontal.b*y] = v;
|
|
}
|
|
}
|
|
}
|
|
|
|
for(auto & b : blocks) b.push();
|
|
bsp::push_reg(hfinish.data(), hfinish.size());
|
|
bsp::push_reg(vfinish.data(), vfinish.size());
|
|
bsp::sync();
|
|
|
|
double time1 = bsp::time();
|
|
|
|
if(true){ // horizontal
|
|
// do local wavelets and communications
|
|
unsigned int stride = 1;
|
|
for(unsigned int i = 0; i < horizontal.big_steps; ++i){
|
|
hstep(d, plan, blocks, stride);
|
|
stride <<= plan.horizontal.m;
|
|
}
|
|
|
|
// finish parallely
|
|
unsigned int hh = horizontal.b/2;
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int t = (d.s - b + d.p)%d.p;
|
|
unsigned int x_start = b * 2;
|
|
auto ptr = blocks[b].data.data();
|
|
for(unsigned int y = 0; y < vertical.b; ++y){
|
|
bsp::put(t, &ptr[0 + horizontal.b*y], hfinish.data(), 0 + x_start + 2*d.p*y, 1);
|
|
bsp::put(t, &ptr[hh + horizontal.b*y], hfinish.data(), 1 + x_start + 2*d.p*y, 1);
|
|
}
|
|
}
|
|
bsp::sync();
|
|
|
|
for(unsigned int y = 0; y < vertical.b; ++y){
|
|
wvlt::wavelet(hfinish.data() + 2*d.p*y, 2*d.p, 1);
|
|
}
|
|
|
|
for(unsigned int y = 0; y < vertical.b; ++y){
|
|
for(unsigned int t = 0; t < d.p; ++t){
|
|
unsigned int b = (t - d.s + d.p)%d.p;
|
|
unsigned int x_start = b * 2;
|
|
bsp::put(t, &hfinish[0 + x_start + 2*d.p*y], blocks[b].data.data(), 0 + horizontal.b*y, 1);
|
|
bsp::put(t, &hfinish[1 + x_start + 2*d.p*y], blocks[b].data.data(), hh + horizontal.b*y, 1);
|
|
}
|
|
}
|
|
bsp::sync();
|
|
}
|
|
|
|
if(true){ //vertical
|
|
// do local wavelets and communications in other directions
|
|
unsigned int stride = 1;
|
|
for(unsigned int i = 0; i < vertical.big_steps; ++i){
|
|
vstep(d, plan, blocks, stride);
|
|
stride <<= plan.vertical.m;
|
|
}
|
|
|
|
// finish parallely
|
|
unsigned int hh = vertical.b/2;
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int t = b;
|
|
unsigned int y_start = (d.s - b + d.p)%d.p * 2;
|
|
auto ptr = blocks[b].data.data();
|
|
for(unsigned int x = 0; x < horizontal.b; ++x){
|
|
bsp::put(t, &ptr[x + 0 *horizontal.b], vfinish.data(), x + horizontal.b*(y_start + 0), 1);
|
|
bsp::put(t, &ptr[x + hh*horizontal.b], vfinish.data(), x + horizontal.b*(y_start + 1), 1);
|
|
}
|
|
}
|
|
bsp::sync();
|
|
|
|
for(unsigned int x = 0; x < horizontal.b; ++x){
|
|
wvlt::wavelet(vfinish.data() + x, 2*d.p, horizontal.b);
|
|
}
|
|
|
|
for(unsigned int x = 0; x < horizontal.b; ++x){
|
|
for(unsigned int t = 0; t < d.p; ++t){
|
|
unsigned int b = d.s;
|
|
unsigned int y_start = (t - b + d.p)%d.p * 2;
|
|
bsp::put(t, &vfinish[x + horizontal.b*(y_start + 0)], blocks[b].data.data(), horizontal.b*0 + x, 1);
|
|
bsp::put(t, &vfinish[x + horizontal.b*(y_start + 1)], blocks[b].data.data(), horizontal.b*hh + x, 1);
|
|
}
|
|
}
|
|
bsp::sync();
|
|
}
|
|
|
|
double time2 = bsp::time();
|
|
if(d.s==0) printf("parallel version\t%f\n", time2 - time1);
|
|
|
|
// Direct write because MCBSP can do this ;D
|
|
for(unsigned int b = 0; b < blocks.size(); ++b){
|
|
unsigned int x_start = b * horizontal.b;
|
|
unsigned int y_start = (d.s - b + d.p)%d.p * vertical.b;
|
|
for(unsigned int y = 0; y < vertical.b; ++y){
|
|
for(unsigned int x = 0; x < horizontal.b; ++x){
|
|
auto v = blocks[b].data[x + horizontal.b*y];
|
|
parr[x_start + x + horizontal.n*(y_start+y)] = v;
|
|
}
|
|
}
|
|
}
|
|
|
|
bsp::end();
|
|
}
|
|
|
|
static void seq_wavelet_2D(){
|
|
seqr = data;
|
|
auto time1 = timer::clock::now();
|
|
wvlt::wavelet_2D(seqr.data(), W, H);
|
|
auto time2 = timer::clock::now();
|
|
printf("sequential version\t%f\n", timer::from_dur(time2 - time1));
|
|
}
|
|
|
|
static double sq_diff(double x, double y){ return (x-y)*(x-y); }
|
|
static int compare_results(std::vector<double> const & lh, std::vector<double> const & rh, double threshold){
|
|
if(lh == rh){
|
|
std::cout << colors::green("SUCCES:") << " bitwise qual" << std::endl;
|
|
return 0;
|
|
}
|
|
|
|
double rmse = std::sqrt(std::inner_product(lh.begin(), lh.end(), rh.begin(), 0.0, std::plus<double>(), &sq_diff) / lh.size());
|
|
if(rmse <= threshold){
|
|
std::cout << colors::green("SUCCES:") << " error within threshold, rmse = " << rmse << std::endl;
|
|
return 1;
|
|
} else {
|
|
std::cout << colors::red("FAIL:") << " error to big, rmse = " << rmse << std::endl;
|
|
return 2;
|
|
}
|
|
}
|
|
|
|
int main(int argc, char** argv){
|
|
P = 2;
|
|
H = 1024;
|
|
W = 1024;
|
|
|
|
data.assign(W*H, 0.0);
|
|
for(unsigned int y = 0; y < H; ++y)
|
|
for(unsigned int x = 0; x < W; ++x)
|
|
data[x + W*y] = x*y;
|
|
seqr.assign(W*H, 0.0);
|
|
parr.assign(W*H, 0.0);
|
|
|
|
bsp::init(par_wavelet_2D, argc, argv);
|
|
par_wavelet_2D();
|
|
seq_wavelet_2D();
|
|
|
|
double threshold = 1.0e-8;
|
|
std::cout << "Checking results ";
|
|
compare_results(seqr, parr, threshold);
|
|
}
|
|
|