diff --git a/assignments/01/Makefile b/assignments/01/Makefile
index 3fe2f05..89bcfa6 100644
--- a/assignments/01/Makefile
+++ b/assignments/01/Makefile
@@ -3,8 +3,8 @@
 CFLAGS := -std=c11 -fopenmp \
 	-I/opt/homebrew/opt/libomp/include \
 	-I/usr/lib/gcc/aarch64-linux-gnu/11/include \
-	-O3
-LDFLAGS :=  -std=c11 -fopenmp -L/opt/homebrew/opt/libomp/lib -O3
+	-O3 -g
+LDFLAGS := -std=c11 -fopenmp -L/opt/homebrew/opt/libomp/lib -O3
 RUST_SOURCES := $(shell find . -name "*.rs")
 
 all: lc_openmp lc_pthreads
@@ -23,7 +23,7 @@ zhan4854.tar.gz: common.c common.h lc_openmp.c lc_pthreads.c Makefile
 	rm -r zhan4854
 
 lc_openmp: lc_openmp.o common.o
-	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^
+	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^ -lm
 
 lc_pthreads: lc_pthreads.o common.o
 	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^
@@ -46,7 +46,7 @@ run-openmp-small: lc_openmp dataset/small/train_data.txt
 	./lc_openmp dataset/small/train_data.txt dataset/small/train_label.txt 10 2 dataset/small/test_data.txt dataset/small/test_label.txt
 
 run-openmp-mnist: lc_openmp dataset/mnist/train_data.txt
-	./lc_openmp dataset/mnist/train_data.txt dataset/mnist/train_label.txt 10 2 dataset/mnist/test_data.txt dataset/mnist/test_label.txt
+	./lc_openmp dataset/mnist/train_data.txt dataset/mnist/train_label.txt 10 8 dataset/mnist/test_data.txt dataset/mnist/test_label.txt
 
 rust: $(RUST_SOURCES)
 	cargo run -- ${BASE_PATH}/dataset/{small_data.csv,small_label.csv} 10 2
diff --git a/assignments/01/lc_openmp.c b/assignments/01/lc_openmp.c
index 577b04e..35e9371 100644
--- a/assignments/01/lc_openmp.c
+++ b/assignments/01/lc_openmp.c
@@ -31,13 +31,16 @@ int main(int argc, char **argv) {
 
   printf("Running %d iteration(s) with %d thread(s).\n", outer_iterations,
          thread_count);
+  double program_start_time = monotonic_seconds();
 
   FLOAT *w = calloc(data->dimensions, sizeof(FLOAT));
   FLOAT *new_w = calloc(data->dimensions, sizeof(FLOAT));
   FLOAT *ouais = calloc(data->dimensions * data->rows, sizeof(FLOAT));
+  // FLOAT *loss_matrix = calloc(data->rows, sizeof(FLOAT));
 
+  double total_compute_time = 0;
   for (int iter = 0; iter < outer_iterations; iter++) {
-    double start_time = monotonic_seconds();
+    double iter_start_time = monotonic_seconds();
 
 #pragma omp parallel for default(shared)
     for (int i = 0; i < data->dimensions; i++) {
@@ -66,26 +69,49 @@ int main(int argc, char **argv) {
         denom += xij * xij;
       }
 
-      if (denom == 0) {
+      if (denom == 0)
         new_w[i] = 0;
-        // printf("wtf? %f\n", numer);
-      } else
+      else
         new_w[i] = numer / denom;
     }
 
-    double end_time = monotonic_seconds();
-    print_time(end_time - start_time);
+    double iter_end_time = monotonic_seconds();
+    total_compute_time += iter_end_time - iter_start_time;
+    printf("Iter duration (no print): %0.04fs\n",
+           iter_end_time - iter_start_time);
 
-    printf("Done.\nw = [");
+    // Update w
+    // printf("w = [");
     for (int idx = 0; idx < data->dimensions; idx++) {
       w[idx] = new_w[idx];
-      printf("%.3f ", w[idx]);
+      // printf("%.3f ", w[idx]);
     }
-    printf("]\n");
+    // printf("]\n");
+
+    // Compute loss
+    FLOAT loss_sum = 0;
+#pragma omp parallel for default(shared)
+    for (int j = 0; j < data->rows; j++) {
+      FLOAT loss_value = 0;
+      for (int i = 0; i < data->dimensions; i++) {
+        loss_value += data->buf[data->rows * i + j] * w[i];
+      }
+
+      loss_value -= label->buf[j];
+      loss_sum += loss_value * loss_value;
+    }
+    FLOAT loss = sqrt(loss_sum);
+    printf("Loss: %0.04f\n", loss);
 
     // memcpy(w, new_w, data->dimensions * sizeof(FLOAT));
   }
 
+  double program_end_time = monotonic_seconds();
+  printf("Program time (compute): %0.04fs\n", total_compute_time);
+  printf("Program time (total): %0.04fs\n",
+         program_end_time - program_start_time);
+
+  // free(loss_matrix);
   free(ouais);
   free(new_w);
   free(data->buf);