From 36a4e033482b45539a709ab55ad64952190b6b33 Mon Sep 17 00:00:00 2001
From: Pranav Sivaraman <14294205+pranav-sivaraman@users.noreply.github.com>
Date: Sun, 5 Nov 2023 13:59:27 -0500
Subject: [PATCH] fix: allocation of host verification vector

---
 cuda/GridInit.cu   | 7 +++++++
 cuda/Simulation.cu | 7 ++-----
 2 files changed, 9 insertions(+), 5 deletions(-)

diff --git a/cuda/GridInit.cu b/cuda/GridInit.cu
index 13c62db..6eb4d53 100644
--- a/cuda/GridInit.cu
+++ b/cuda/GridInit.cu
@@ -100,6 +100,8 @@ SimulationData grid_init_do_not_profile( Inputs in, int mype )
 {
 	// Structure to hold all allocated simuluation data arrays
 	SimulationData SD;
+	
+
 
 	// Keep track of how much data we're allocating
 	size_t nbytes = 0;
@@ -152,6 +154,11 @@ SimulationData grid_init_do_not_profile( Inputs in, int mype )
 			printf("E%d = %lf\n", j, SD.nuclide_grid[i * in.n_gridpoints + j].energy);
 	}
 	*/
+
+    // Allocate Verification Array
+    size_t sz = in.lookups * sizeof(unsigned long);
+    SD.verification = (unsigned long *) malloc(sz);
+    nbytes += sz;
 	
 
 	////////////////////////////////////////////////////////////////////
diff --git a/cuda/Simulation.cu b/cuda/Simulation.cu
index 8be81ab..c7b8a73 100644
--- a/cuda/Simulation.cu
+++ b/cuda/Simulation.cu
@@ -14,9 +14,6 @@
 
 unsigned long long run_event_based_simulation_baseline(Inputs in, SimulationData SD, int mype)
 {
-    size_t sz = in.lookups * sizeof(unsigned long);
-    unsigned long * v = (unsigned long *) malloc(sz);
-    
     // Move Data to GPU
     SimulationData GSD = move_simulation_data_to_device(in, mype, SD);
 
@@ -36,11 +33,11 @@ unsigned long long run_event_based_simulation_baseline(Inputs in, SimulationData
 	// Reduce Verification Results
 	////////////////////////////////////////////////////////////////////////////////
 	if( mype == 0)	printf("Reducing verification results...\n");
-    gpuErrchk(cudaMemcpy(v, GSD.verification, sz, cudaMemcpyDeviceToHost) );
+    gpuErrchk(cudaMemcpy(SD.verification, GSD.verification, in.lookups * sizeof(unsigned long), cudaMemcpyDeviceToHost) );
 
     unsigned long verification_scalar = 0;
     for( int i =0; i < in.lookups; i++ )
-        verification_scalar += v[i];
+        verification_scalar += SD.verification[i];
 
     release_device_memory(GSD);