Modified mp functions using OpenMP and OpenBLAS

Author: audictheprogrammer

src/data/benchmark_blocks_BLAS.txt

@@ -1,3 +1,3 @@
-256 0.111944 
-512 1.544904 
-1024 11.227543 
+256 0.106118 
+512 2.280422 
+1024 10.936971 

src/data/benchmark_blocks_BLAS_MP.txt

@@ -1,3 +1,3 @@
-256 0.053706 
-512 6.166882 
-1024 16.398344 
+256 0.133424 
+512 0.738715 
+1024 4.116122 

src/data/benchmark_blocks_NoBLAS.txt

@@ -1,3 +1,3 @@
-256 0.376757 
-512 2.228056 
-1024 13.547425 
+256 0.729653 
+512 1.982209 
+1024 15.146636 

src/data/benchmark_modulo_MP_Barrett.txt

@@ -1,3 +1,3 @@
-256 0.061757 
-512 0.533625 
-1024 3.274027 
+256 0.060192 
+512 0.509851 
+1024 3.235947 

src/data/benchmark_modulo_MP_SIMD1.txt

@@ -1,3 +1,3 @@
-256 0.104602 
-512 0.478607 
-1024 3.103020 
+256 0.055740 
+512 0.544654 
+1024 4.393463 

src/data/benchmark_modulo_MP_SIMD2.txt

@@ -1,3 +1,3 @@
-256 0.360917 
-512 0.988932 
-1024 4.843099 
+256 0.054218 
+512 0.442351 
+1024 3.072420 

src/data/benchmark_modulo_MP_SIMD3.txt

@@ -1,3 +1,3 @@
-256 0.173421 
-512 0.971074 
-1024 4.871781 
+256 0.090017 
+512 0.436566 
+1024 2.908950 

src/data/benchmark_modulo_MP_naive.txt

@@ -1,3 +1,3 @@
-256 0.110528 
-512 0.436584 
-1024 2.869017 
+256 0.090896 
+512 0.444606 
+1024 3.990551 

src/main

@@ -270,7 +270,7 @@ void benchmark_loops_order(double p){
     /* Benchmarking the order of loops.
     The most efficient one is IKJ.
     */
-    int m = 3;  // Executes m times each algo
+    int m = 1;  // Executes m times each algo
     for (int i=8; i<11; i++){
         int n = (int) pow(2, i);
         double sum_ijk = 0;
@@ -308,7 +308,7 @@ void benchmark_modulos(double p, double u, double u_overline, double u_b){
     /* Benchmarking different modulos.
     The most efficient one SIMD2.
     */
-    int m = 3;  // Executes m times each algo
+    int m = 1;  // Executes m times each algo
     for (int i=8; i<11; i++){
         int n = (int) pow(2, i);
         double sum_mod_naive = 0;
@@ -344,7 +344,7 @@ void benchmark_modulos_MP(double p, double u, double u_overline, double u_b){
     /* Benchmarking different modulos.
     The most efficient one SIMD2.
     */
-    int m = 3;  // Executes m times each algo
+    int m = 1;  // Executes m times each algo
     for (int i=8; i<11; i++){
         int n = (int) pow(2, i);
 
@@ -381,7 +381,7 @@ void benchmark_modulos_MP(double p, double u, double u_overline, double u_b){
 void benchmark_blocks(double p, double u_overline){
     /* Benchmarking different modulos.
     */
-    int m = 3;  // Executes m times each algo
+    int m = 1;  // Executes m times each algo
     for (int i=8; i<11; i++){
         int n = (int) pow(2, i);
         int b = get_blocksize(get_bitsize(p), n);
@@ -508,15 +508,15 @@ int main(){
     // benchmark_modulos(p, u, u_overline, u_b);
 
     // Benchmarking different modulos with OpenMP
-    clean_file_modulos_MP();
-    benchmark_modulos_MP(p, u, u_overline, u_b);
+    // clean_file_modulos_MP();
+    // benchmark_modulos_MP(p, u, u_overline, u_b);
 
     // Benchmarking blocks.
-    // clean_file_blocks();
-    // benchmark_blocks(p, u_overline);
+    clean_file_blocks();
+    benchmark_blocks(p, u_overline);
 
-    clean_file_blocks_MP();
-    benchmark_blocks_MP(p, u_overline);
+    // clean_file_blocks_MP();
+    // benchmark_blocks_MP(p, u_overline);
 
 
     return 0;

src/main.c

@@ -199,8 +199,8 @@ void mp_SIMD1_MP(double* A, double* B, double* C, int n, double p, double u){
 void mp_SIMD2_MP(double* A, double* B, double* C, int n, double p, double u){
     // Assert C is a zero matrix
     for (int k=0; k<n; k++){
+        #pragma omp parallel for
         for (int i=0; i<n; i++){
-            #pragma omp parallel for
             for (int j=0; j<n; j++){
                     C[i*n + j] = C[i*n + j] + modulo_SIMD2(A[i*n + k] * B[k*n + j], p, u);
             }
@@ -274,10 +274,8 @@ void mp_block(double* A, double* B, double* C, int n, double p, double u, int b)
             }
         }
 
-        for (int i=0; i<n; i++){
-            for (int j=0; j<n; j++){
-                C[i*n + j] = modulo_SIMD3(C[i*n + j], p, u);
-            }
+        for (int i=0; i<n*n; i++){
+            C[i] = modulo_SIMD3(C[i], p, u);
         }
 
     }
@@ -288,16 +286,15 @@ void mp_block_BLAS(double* A, double* B, double* C, int n, double p, double u, i
     /* Compute the product of two matrices using OpenBLAS's block product.
     It allows to reduce the amount of modulo needed.
     */
-
+    openblas_set_num_threads(1);  // 8 is slower than 1.
     for (int k=0; k<n; k+=b){
         cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
                     n, n, b, 1, A + k, n, B + n*k, n,
                     1, C, n);
 
-        for (int i=0; i<n; i++){
-            for (int j=0; j<n; j++){
-                C[i*n + j] = modulo_SIMD3(C[i*n + j], p, u);
-            }
+
+        for (int i=0; i<n*n; i++){
+            C[i] = modulo_SIMD3(C[i], p, u);
         }
 
     }
@@ -306,19 +303,16 @@ void mp_block_BLAS(double* A, double* B, double* C, int n, double p, double u, i
 
 void mp_block_BLAS_MP(double* A, double* B, double* C, int n, double p, double u, int b){
     /* Compute the product of two matrices using OpenBLAS's block product.
-    It allows to reduce the amount of modulo needed.
+    It allows us to reduce the amount of modulo needed.
     */
-
     for (int k=0; k<n; k+=b){
         cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
                     n, n, b, 1, A + k, n, B + n*k, n,
                     1, C, n);
 
         #pragma omp parallel for
-        for (int i=0; i<n; i++){
-            for (int j=0; j<n; j++){
-                C[i*n + j] = modulo_SIMD3(C[i*n + j], p, u);
-            }
+        for (int i=0; i<n*n; i++){
+            C[i] = modulo_SIMD3(C[i], p, u);
         }
 
     }

src/main_blas

@@ -15,9 +15,9 @@ set style line 6 lc rgb '#ff00ff' lt 1 lw 2 pt 7 pi -1 ps 1.5
 set pointintervalbox 3
 
 plot "../data/benchmark_modulo_SIMD2.txt" using 1:2 title "No blocks" with linespoints ls 1
-replot "../data/benchmark_modulo_MP_SIMD2.txt" using 1:2 title "MP SIMD2" with linespoints ls 2
+replot "../data/benchmark_modulo_MP_SIMD2.txt" using 1:2 title "MP SIMD3" with linespoints ls 2
 replot "../data/benchmark_blocks_BLAS.txt" using 1:2 title "BLAS" with linespoints ls 3
 replot "../data/benchmark_blocks_BLAS_MP.txt" using 1:2 title "BLAS MP" with linespoints ls 4
-replot "../data/benchmark_order_kij.txt" using 1:2 title "No Mod" with linespoints ls 5
+# replot "../data/benchmark_order_kij.txt" using 1:2 title "No Mod" with linespoints ls 5
 
 # Graph Blocks product with multithreads

src/main_test

@@ -20,11 +20,12 @@ set pointintervalbox 3
 # replot "../data/benchmark_modulo_MP_SIMD1.txt" using 1:2 title "MP SIMD1" with linespoints ls 3
 # replot "../data/benchmark_modulo_SIMD1.txt" using 1:2 title "SIMD1" with linespoints ls 3
 
-plot "../data/benchmark_modulo_MP_SIMD2.txt" using 1:2 title "MP SIMD2" with linespoints ls 3
-replot "../data/benchmark_modulo_SIMD2.txt" using 1:2 title "SIMD2" with linespoints ls 4
+plot "../data/benchmark_modulo_SIMD2.txt" using 1:2 title "SIMD2" with linespoints ls 3
+replot "../data/benchmark_modulo_SIMD3.txt" using 1:2 title "SIMD3" with linespoints ls 4
+
+replot "../data/benchmark_modulo_MP_SIMD2.txt" using 1:2 title "MP SIMD2" with linespoints ls 5
+replot "../data/benchmark_modulo_MP_SIMD3.txt" using 1:2 title "MP SIMD3" with linespoints ls 6
 
-replot "../data/benchmark_modulo_MP_SIMD3.txt" using 1:2 title "MP SIMD3" with linespoints ls 5
-replot "../data/benchmark_modulo_SIMD3.txt" using 1:2 title "SIMD3" with linespoints ls 6
 
 # replot "../data/benchmark_modulo_MP_Barrett.txt" using 1:2 title "MPBarrett" with linespoints ls 6
 # replot "../data/benchmark_modulo_Barrett.txt" using 1:2 title "Barrett" with linespoints ls 6