diff --git a/wavelet/jcmp.cpp b/wavelet/jcmp.cpp
index b768350..183c772 100644
--- a/wavelet/jcmp.cpp
+++ b/wavelet/jcmp.cpp
@@ -109,14 +109,15 @@ int main(){
 
 		// compress and decompress to see how we lost information
 		unsigned int nzeros = 0;
-		auto compressed_vec = decompress(compress(image_vec, width, 0.5, nzeros));
+		auto compressed_vec = decompress(compress(image_vec, width, 5.0, nzeros));
 
 		// output some information
-		std::cout << field("raw") << human_string(sizeof(uint8_t) * image_vec.size(), "b") << std::endl;
-		std::cout << field("compressed") << human_string(sizeof(jcmp::coefficient) * nzeros, "b") << std::endl;
+		std::cout << field("raw") << width*height << std::endl;
+		std::cout << field("compressed") << nzeros/double(width*height) << std::endl;
+		std::cout << field("nz") << nzeros << std::endl;
 
 		// save to output file
-		std::string cfilename = "compressed/" + path.filename().string();
+		std::string cfilename = "compressed/" + std::to_string(nzeros) + path.filename().string();
 		png::gray_ostream compressed_image(width, height, cfilename);
 		for(unsigned int i = 0; i < compressed_vec.size(); ++i) compressed_image << compressed_vec[i];
 	}
diff --git a/wavelet/wavelet_2.hpp b/wavelet/wavelet_2.hpp
index 52e7a1c..2772519 100644
--- a/wavelet/wavelet_2.hpp
+++ b/wavelet/wavelet_2.hpp
@@ -42,8 +42,9 @@ namespace wvlt{
 		// x1 and x2 are next elements, or wrap around
 		// calculates size/stride elements of the output
 		inline void wavelet_mul(double* x, double x1, double x2, unsigned int size, unsigned int stride){
-			assert(x && is_even(size) && is_pow_of_two(stride) && 4*stride <= size);
-			wavelet_mul_base(x, size, stride);
+			assert(x && is_even(size) && is_pow_of_two(stride) && 2*stride <= size);
+			if(4*stride <= size)
+				wavelet_mul_base(x, size, stride);
 
 			unsigned int i = size - 2*stride;
 			double y1 = x[i] * evn_coef[0] + x[i+stride] * evn_coef[1] + x1 * evn_coef[2] + x2 * evn_coef[3];
diff --git a/wavelet/wavelet_parallel.hpp b/wavelet/wavelet_parallel.hpp
index 059be6f..6544405 100644
--- a/wavelet/wavelet_parallel.hpp
+++ b/wavelet/wavelet_parallel.hpp
@@ -36,7 +36,7 @@ namespace wvlt {
 			unsigned int n, b, m, Cm, small_steps, big_steps, remainder;
 
 			plan_1D(unsigned int n_, unsigned int b_, unsigned int m_)
-			: n(n_), b(b_), m(m_), Cm(pow_two(m+1) - 2), small_steps(two_log(b) - 1), big_steps(small_steps/m), remainder(small_steps - m*big_steps)
+			: n(n_), b(b_), m(m_), Cm(pow_two(m+1) - 2), small_steps(two_log(b)), big_steps((small_steps-1)/m), remainder(small_steps - m*big_steps)
 			{}
 		};
 
@@ -81,20 +81,19 @@ namespace wvlt {
 			// First do the local part
 			base(d, plan, x, next, plan.b);
 
+			// we only have to finish centralized if p >= 4
+			if(d.p <= 2) return;
+
 			// Then do a fan in (i.e. 2 elements to proc zero)
-			for(unsigned int i = 0; i < 2; ++i){
-				bsp::put(0, &x[i * plan.b/2], proczero, d.s * 2 + i);
-			}
+			bsp::put(0, x, proczero, d.s);
 			bsp::sync();
 
 			// proc zero has the privilige/duty to finish the job
 			if(d.s == 0) {
-				wvlt::wavelet(proczero, 2*d.p, 1);
+				wvlt::wavelet(proczero, d.p, 1);
 				// and to send it back to everyone
 				for(unsigned int t = 0; t < d.p; ++t){
-					for(unsigned int i = 0; i < 2; ++i){
-						bsp::put(t, &proczero[t*2 + i], x, i * plan.b/2);
-					}
+					bsp::put(t, &proczero[t], x);
 				}
 			}
 		}
diff --git a/wavelet/wavelet_parallel_mockup.cpp b/wavelet/wavelet_parallel_mockup.cpp
index 5067e2d..0bc9307 100644
--- a/wavelet/wavelet_parallel_mockup.cpp
+++ b/wavelet/wavelet_parallel_mockup.cpp
@@ -57,7 +57,7 @@ static void par_wavelet(){
 	// For convenience and consistency we use std::vector
 	std::vector<double> x(plan.b, 0.0);
 	std::vector<double> next(plan.Cm, 0.0);
-	std::vector<double> proczero(d.s == 0 ? 2*d.p : 1, 0.0);
+	std::vector<double> proczero(d.s == 0 ? d.p : 1, 0.0);
 
 	bsp::push_reg(x.data(), x.size());
 	bsp::push_reg(next.data(), next.size());
diff --git a/wavelet_report/conc.tex b/wavelet_report/conc.tex
index 22fd73b..37d18c3 100644
--- a/wavelet_report/conc.tex
+++ b/wavelet_report/conc.tex
@@ -12,3 +12,4 @@ As for the application of image compression, we have seen that the wavelets capt
 	\includegraphics[width=0.3\textwidth]{parijs.png}
 	\caption{My girlfriend and I in Paris. Still romantic with only 200 coefficients.}
 \end{figure}
+\vspace{3cm}
diff --git a/wavelet_report/report.tex b/wavelet_report/report.tex
index 9ead65b..79b64e0 100644
--- a/wavelet_report/report.tex
+++ b/wavelet_report/report.tex
@@ -21,7 +21,9 @@ In this paper we will derive a parallel algorithm to perform a Daubechies wavele
 \include{par}
 \include{img}
 \include{res}
-\include{conc}
+
+\newpage
+\input{conc}
 
 
 \nocite{*}