experimental implementation for tc-scanb > tc-aggrj, the second

qdp_project/src/Benchmark.cpp

@@ -39,12 +39,10 @@ uint64_t* DATA_DST_;
 // if more j than b -> subsplit b like it is now
 
 void scan_b(size_t gid, size_t tid) {
-    constexpr size_t RUN_INCREMENT = TC_SCANB / TC_AGGRJ;
-    constexpr size_t THREAD_RATIO = TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
-    constexpr size_t SUBCHUNK_COUNT = THREAD_RATIO > 0 ? THREAD_RATIO : 1;
-    constexpr size_t SUBCHUNK_SIZE_B = CHUNK_SIZE_B / SUBCHUNK_COUNT;
-    constexpr size_t SUBCHUNK_SIZE_ELEMENTS = CHUNK_SIZE_ELEMENTS / SUBCHUNK_COUNT;
-    constexpr size_t TC_ACTUAL = TC_AGGRJ >= TC_SCANB ? TC_SCANB : TC_AGGRJ;
+    constexpr size_t VIRT_TID_INCREMENT = TC_SCANB / TC_AGGRJ;
+    constexpr size_t SUBCHUNK_THREAD_RATIO = TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
+    constexpr bool CACHE_SUBCHUNKING = SUBCHUNK_THREAD_RATIO > 1;
+    constexpr bool CACHE_OVERCHUNKING = VIRT_TID_INCREMENT > 1;
 
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)].clear();
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)].resize(1);
@@ -54,15 +52,38 @@ void scan_b(size_t gid, size_t tid) {
     THREAD_TIMING_[SCANB_TIMING_INDEX][UniqueIndex(gid,tid)][0][TIME_STAMP_BEGIN] = std::chrono::steady_clock::now();
 
     if constexpr (PERFORM_CACHING) {
-        for (size_t i = 0; i < RUN_COUNT; i += RUN_INCREMENT) {
-            const size_t chunk_index = get_chunk_index(gid, i);
-            uint64_t* chunk_ptr = get_chunk<TC_ACTUAL>(DATA_B_, chunk_index, tid);
-
-            for (size_t j = 0; j < SUBCHUNK_COUNT; j++) {
-                uint64_t* sub_chunk_ptr = &chunk_ptr[j * SUBCHUNK_SIZE_ELEMENTS];
-                CACHE_.Access(reinterpret_cast<uint8_t*>(sub_chunk_ptr), SUBCHUNK_SIZE_B);
+        if constexpr (CACHE_SUBCHUNKING) {
+            constexpr size_t SUBCHUNK_COUNT = SUBCHUNK_THREAD_RATIO > 0 ? SUBCHUNK_THREAD_RATIO : 1;
+            constexpr size_t SUBCHUNK_SIZE_B = CHUNK_SIZE_B / SUBCHUNK_COUNT;
+            constexpr size_t SUBCHUNK_SIZE_ELEMENTS = CHUNK_SIZE_ELEMENTS / SUBCHUNK_COUNT;
+
+            for (size_t i = 0; i < RUN_COUNT; i++) {
+                const size_t chunk_index = get_chunk_index(gid, i);
+                uint64_t* chunk_ptr = get_chunk<TC_SCANB>(DATA_B_, chunk_index, tid);
+
+                for (size_t j = 0; j < SUBCHUNK_COUNT; j++) {
+                    uint64_t* sub_chunk_ptr = &chunk_ptr[j * SUBCHUNK_SIZE_ELEMENTS];
+                    CACHE_.Access(reinterpret_cast<uint8_t*>(sub_chunk_ptr), SUBCHUNK_SIZE_B);
+                }
+            }
+        }
+        else if constexpr (CACHE_OVERCHUNKING) {
+            for (size_t tid_virt = tid; tid_virt < TC_AGGRJ; tid_virt += VIRT_TID_INCREMENT) {
+                for (size_t i = 0; i < RUN_COUNT; i++) {
+                    const size_t chunk_index = get_chunk_index(gid, i);
+                    uint64_t* chunk_ptr = get_chunk<TC_AGGRJ>(DATA_B_, chunk_index, tid_virt);
+                    CACHE_.Access(reinterpret_cast<uint8_t*>(chunk_ptr), CHUNK_SIZE_B);
+                }
             }
         }
+        else {
+            for (size_t i = 0; i < RUN_COUNT; i++) {
+                const size_t chunk_index = get_chunk_index(gid, i);
+                uint64_t* chunk_ptr = get_chunk<TC_SCANB>(DATA_B_, chunk_index, tid);
+                CACHE_.Access(reinterpret_cast<uint8_t*>(chunk_ptr), CHUNK_SIZE_B);
+            }
+        }
+
     }
 
     if constexpr (COMPLEX_QUERY) {

qdp_project/src/Configuration.hpp

@@ -79,8 +79,6 @@ constexpr size_t CHUNK_COUNT = WL_SIZE_B / CHUNK_SIZE_B;
 constexpr size_t CHUNK_SIZE_ELEMENTS = CHUNK_SIZE_B / sizeof(uint64_t);
 constexpr size_t RUN_COUNT = CHUNK_COUNT / GROUP_COUNT;
 
-static_assert(TC_AGGRJ % (TC_SCANB > 0 ? TC_SCANB : TC_AGGRJ) == 0);
-static_assert(TC_AGGRJ >= TC_SCANB);
 static_assert(RUN_COUNT > 0);
 static_assert(WL_SIZE_B % 16 == 0);
 static_assert(CHUNK_SIZE_B % 16 == 0);