Merge branch 'master' of https://git.constantin-fuerst.com/constantin/bachelor-thesis

offloading-cacher/cache.hpp

@@ -400,12 +400,6 @@ inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, con
     const size_t size = task->GetSize() / task_count;
     const size_t last_size = size + task->GetSize() % task_count;
 
-    // save the current numa node mask to restore later
-    // as executing the copy task will place this thread
-    // on a different node
-
-    bitmask* nodemask = numa_get_run_node_mask();
-
     auto handlers = new std::vector<dml_job_t*>();
 
     for (uint32_t i = 0; i < task_count; i++) {
@@ -418,18 +412,11 @@ inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, con
     }
 
     task->SetTaskHandlersAndCache(dst, handlers);
-
-    // restore the previous nodemask
-
-    numa_run_on_node_mask(nodemask);
-    numa_free_nodemask(nodemask);
 }
 
 inline dml_job_t* dsacache::Cache::ExecuteCopy(
         const uint8_t* src, uint8_t* dst, const size_t size, const int node
 ) const {
-    numa_run_on_node(node);
-
     uint32_t job_size = 0;
     dml_status_t status = dml_get_job_size(DML_PATH_HW, &job_size);
 
@@ -452,6 +439,7 @@ inline dml_job_t* dsacache::Cache::ExecuteCopy(
     job->destination_first_ptr = dst;
     job->source_length = size;
     job->flags |= DML_FLAG_BLOCK_ON_FAULT | DML_FLAG_COPY_ONLY;
+    job->numa_id = node;
 
     status = dml_submit_job(job);
 
@@ -642,9 +630,13 @@ inline dsacache::CacheData::~CacheData() {
 
         Deallocate();
 
+	std::vector<dml_job_t*>* handlers = handlers_->load();
+
+	if (handlers != nullptr && handlers != reinterpret_cast<std::vector<dml_job_t*>*>(maxptr)) {
 	    for (dml_job_t* job : *handlers_->load()) {
 	        if (job != nullptr) delete job;
 	    }
+        }
 
         delete active_;
         delete cache_;

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

@@ -1,6 +0,0 @@
-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

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;97400868;0;
+1;97565944;0;
+2;89098555;0;
+3;93226925;0;
+4;97550283;0;

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

@@ -1,6 +0,0 @@
-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

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;67853704;0;
+1;85513791;0;
+2;66482278;0;
+3;67492755;0;
+4;68083298;0;

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

@@ -1,6 +0,0 @@
-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-tca1-tcb1-tcj1-tmul32-wl4294967296-cs134217728.csv

@@ -0,0 +1,6 @@
+run;time;result[0];
+0;6136464327;0;
+1;6125809164;0;
+2;6122382339;0;
+3;6207611081;0;
+4;6149495534;0;

qdp_project/src/Benchmark.cpp

@@ -20,7 +20,7 @@
 
 #include "BenchmarkHelpers.cpp"
 
-#define MODE_PREFETCH
+#define MODE_HBM
 
 ////////////////////////////////
 /// BENCHMARK SETUP
@@ -30,8 +30,8 @@ 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 = 8;
+constexpr size_t CHUNK_SIZE_B = 128_MiB;
+constexpr uint32_t GROUP_COUNT = 32;
 constexpr uint32_t TC_SCANA = 1;
 constexpr uint32_t TC_SCANB = 2;
 constexpr uint32_t TC_AGGRJ = 2;
@@ -74,7 +74,6 @@ constexpr size_t MASK_STEP_SIZE = CHUNK_SIZE_ELEMENTS / MASK_ELEMENT_SIZE;
 
 static_assert(RUN_COUNT > 0);
 static_assert(TC_SCANB <= TC_AGGRJ);
-static_assert(TC_AGGRJ % TC_SCANB == 0);
 static_assert(WL_SIZE_B % 16 == 0);
 static_assert(CHUNK_SIZE_B % 16 == 0);
 
@@ -168,7 +167,7 @@ void process_timings(
 }
 
 void scan_b(size_t gid, size_t tid) {
-    constexpr size_t split =  TC_AGGRJ / TC_SCANB;
+    constexpr size_t split =  TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
     const size_t start = tid * split;
     const size_t end = start + split;