add prelimianry results, modify the launch script and provide macro-based selection of parameters for the three modes {dram,hbm,prefetch}

qdp_project/bench_max.sh

@@ -3,7 +3,7 @@
 current_date_time=$(date)
 echo "Benchmark start at: $current_date_time"
 
-sudo numactl --cpunodebind=0 cmake-build-release/MAXBench
+sudo numactl --cpunodebind=0 cmake-build-release/QDPBench
 
 current_date_time=$(date)
 echo "Benchmark end at: $current_date_time"

qdp_project/results/qdp-xeonmax-simpleq-DramBase-tca4-tcb0-tcj2-tmul4-wl4294967296-cs1048576.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+5;72414484;0;
+6;72695538;0;
+7;70226208;0;
+8;64734077;0;
+9;70125966;0;

qdp_project/results/qdp-xeonmax-simpleq-HbmPeak-tca4-tcb0-tcj2-tmul4-wl4294967296-cs1048576.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+5;67105366;0;
+6;64812778;0;
+7;64667008;0;
+8;66072639;0;
+9;66492940;0;

qdp_project/results/qdp-xeonmax-simpleq-Prefetch-tca4-tcb1-tcj2-tmul4-wl4294967296-cs134217728.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+5;4700269615;0;
+6;4780414439;0;
+7;4596210467;0;
+8;4679035690;0;
+9;4687101402;0;

qdp_project/src/Benchmark.cpp

@@ -19,25 +19,50 @@
 
 #include "BenchmarkHelpers.cpp"
 
+#define MODE_PREFETCH
+
 ////////////////////////////////
 /// BENCHMARK SETUP
 
-constexpr size_t WL_SIZE_B = 64_MiB;
-constexpr size_t CHUNK_SIZE_B = 1_MiB;
-constexpr uint64_t CMP_A = 50;
-constexpr uint32_t WARMUP_ITERATION_COUNT = 0;
-constexpr uint32_t ITERATION_COUNT = 2;
-constexpr uint32_t GROUP_COUNT = 2;
+constexpr size_t WL_SIZE_B = 4_GiB;
+constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
+constexpr uint32_t ITERATION_COUNT = 5;
+
+#ifdef MODE_PREFETCH
+constexpr size_t CHUNK_SIZE_B = 128_MiB;
+constexpr uint32_t GROUP_COUNT = 4;
+constexpr uint32_t TC_SCANA = 4;
 constexpr uint32_t TC_SCANB = 1;
-constexpr uint32_t TC_SCANA = 1;
-constexpr uint32_t TC_AGGRJ = 1;
+constexpr uint32_t TC_AGGRJ = 2;
 constexpr bool PERFORM_CACHING = true;
 constexpr bool DATA_IN_HBM = false;
-constexpr char MODE_STRING[] = "DramBase";
+constexpr char MODE_STRING[] = "Dram";
+#endif
+#ifdef MODE_DRAM
+constexpr size_t CHUNK_SIZE_B = 2_MiB;
+constexpr uint32_t GROUP_COUNT = 4;
+constexpr uint32_t TC_SCANA = 4;
+constexpr uint32_t TC_SCANB = 0;
+constexpr uint32_t TC_AGGRJ = 2;
+constexpr bool PERFORM_CACHING = false;
+constexpr bool DATA_IN_HBM = false;
+constexpr char MODE_STRING[] = "Prefetch";
+#endif
+#ifdef MODE_HBM
+constexpr size_t CHUNK_SIZE_B = 2_MiB;
+constexpr uint32_t GROUP_COUNT = 4;
+constexpr uint32_t TC_SCANA = 4;
+constexpr uint32_t TC_SCANB = 0;
+constexpr uint32_t TC_AGGRJ = 2;
+constexpr bool PERFORM_CACHING = false;
+constexpr bool DATA_IN_HBM = true;
+constexpr char MODE_STRING[] = "Hbm";
+#endif
 
 /// DO NOT CONFIGURE BEYOND THIS
 ////////////////////////////////
 
+constexpr uint64_t CMP_A = 50;
 constexpr uint32_t TC_COMBINED = TC_SCANA + TC_SCANB + TC_AGGRJ;
 constexpr size_t WL_SIZE_ELEMENTS = WL_SIZE_B / sizeof(uint64_t);
 constexpr size_t CHUNK_COUNT = WL_SIZE_B / CHUNK_SIZE_B;
@@ -208,4 +233,4 @@ int main() {
     numa_free(DATA_DST_, TC_AGGRJ * GROUP_COUNT * sizeof(uint64_t));
 
     return 0;
-}
+}