transfer commit from crobat

qdp_project/results/qdp-xeonmax-simpleq-dram-tca4-tcb0-tcj2-tmul4-wl2147483648-cs2097152.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;22199017;0;
+1;16588422;0;
+2;18267635;0;
+3;17026004;0;
+4;16958071;0;

qdp_project/results/qdp-xeonmax-simpleq-dram-tca4-tcb0-tcj2-tmul4-wl4294967296-cs2097152.csv

@@ -1,6 +0,0 @@
-run;time;result[0];
-0;67578888;0;
-1;73660237;0;
-2;73616193;0;
-3;76092000;0;
-4;70963091;0;

qdp_project/results/qdp-xeonmax-simpleq-hbm-tca4-tcb0-tcj2-tmul4-wl2147483648-cs2097152.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;14448722;0;
+1;17734795;0;
+2;19240141;0;
+3;15579654;0;
+4;14252101;0;

qdp_project/results/qdp-xeonmax-simpleq-hbm-tca4-tcb0-tcj2-tmul4-wl4294967296-cs2097152.csv

@@ -1,6 +0,0 @@
-run;time;result[0];
-0;59248029;0;
-1;64829305;0;
-2;65696823;0;
-3;59917878;0;
-4;64575991;0;

qdp_project/results/qdp-xeonmax-simpleq-prefetch-tca1-tcb1-tcj1-tmul32-wl2147483648-cs134217728.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;916843;0;
+1;1060229;0;
+2;914006;0;
+3;1217119;0;
+4;1029607;0;

qdp_project/results/qdp-xeonmax-simpleq-prefetch-tca2-tcb1-tcj4-tmul4-wl4294967296-cs268435456.csv

@@ -1,6 +0,0 @@
-run;time;result[0];
-0;6977961232;0;
-1;7020570076;0;
-2;6935229235;0;
-3;7109847364;0;
-4;6920307956;0;

qdp_project/results/qdp-xeonmax-simpleq-prefetch-tca4-tcb1-tcj4-tmul4-wl4294967296-cs268435456.csv

@@ -1,6 +0,0 @@
-run;time;result[0];
-0;7233530125;0;
-1;7086960403;0;
-2;6962089706;0;
-3;6915175958;0;
-4;6964562288;0;

qdp_project/src/Benchmark.cpp

@@ -19,21 +19,21 @@
 
 #include "BenchmarkHelpers.cpp"
 
-#define MODE_HBM
+#define MODE_PREFETCH
 
 ////////////////////////////////
 /// BENCHMARK SETUP
 
-constexpr size_t WL_SIZE_B = 4_GiB;
+constexpr size_t WL_SIZE_B = 2_GiB;
 constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
 constexpr uint32_t ITERATION_COUNT = 5;
 
 #ifdef MODE_PREFETCH
-constexpr size_t CHUNK_SIZE_B = 256_MiB;
-constexpr uint32_t GROUP_COUNT = 4;
-constexpr uint32_t TC_SCANA = 4;
+constexpr size_t CHUNK_SIZE_B = 128_MiB;
+constexpr uint32_t GROUP_COUNT = ;
+constexpr uint32_t TC_SCANA = 1;
 constexpr uint32_t TC_SCANB = 1;
-constexpr uint32_t TC_AGGRJ = 4;
+constexpr uint32_t TC_AGGRJ = 1;
 constexpr bool PERFORM_CACHING = true;
 constexpr bool DATA_IN_HBM = false;
 constexpr char MODE_STRING[] = "prefetch";
@@ -111,6 +111,8 @@ inline uint16_t* get_mask(uint16_t* base, const size_t chunk_index) {
 }
 
 void scan_b(size_t gid, size_t tid) {
+    using namespace std::chrono_literals;
+
     LAUNCH_.wait();
 
     if constexpr (PERFORM_CACHING) {
@@ -121,11 +123,12 @@ void scan_b(size_t gid, size_t tid) {
             uint64_t* chunk_ptr = get_chunk(DATA_B_, chunk_index);
 
             data = CACHE_.Access(reinterpret_cast<uint8_t *>(chunk_ptr), CHUNK_SIZE_B / TC_AGGRJ);
+    	    std::this_thread::sleep_for(100ms);
+	    data->WaitOnCompletion();
         }
-
-        data->WaitOnCompletion();
     }
 	
+
     BARRIERS_[gid]->arrive_and_drop();
 }