use temporary accumulator started from zero for more precision (#3318)

david-cortes-intel · web-flow · commit 8dbb081fcd12 · 2025-08-07T16:04:47.000+01:00
diff --git a/cpp/daal/src/algorithms/kmeans/kmeans_lloyd_impl.i b/cpp/daal/src/algorithms/kmeans/kmeans_lloyd_impl.i
@@ -302,7 +302,8 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer
         SpBlasInst<algorithmFPType, cpu>::xxcsrmm(&transa, &_n, &_c, &_p, &alpha, matdescra, data, (DAAL_INT *)colIdx, (DAAL_INT *)rowIdx, inClusters,
                                                   &_p, &beta, x_clusters, &_n);
 
-        size_t csrCursor = 0;
+        algorithmFPType goal = 0;
+        size_t csrCursor     = 0;
         for (size_t i = 0; i < blockSize; i++)
         {
             algorithmFPType minGoalVal = clustersSq[0] - x_clusters[i];
@@ -329,7 +330,7 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer
 
             kmeansInsertCandidate(tt, minGoalVal, k * blockSizeDefault + i);
 
-            *trg += minGoalVal;
+            goal += minGoalVal;
 
             cS0[minIdx]++;
 
@@ -339,6 +340,7 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer
                 assignments[i] = (int)minIdx;
             }
         }
+        *trg += goal;
     });
     return safeStat.detach();
 }

Original file line number	Diff line number	Diff line change
`@@ -302,7 +302,8 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer`
`302`	`302`	`SpBlasInst<algorithmFPType, cpu>::xxcsrmm(&transa, &_n, &_c, &_p, &alpha, matdescra, data, (DAAL_INT )colIdx, (DAAL_INT )rowIdx, inClusters,`
`303`	`303`	`&_p, &beta, x_clusters, &_n);`
`304`	`304`
`305`		`- size_t csrCursor = 0;`
	`305`	`+ algorithmFPType goal = 0;`
	`306`	`+ size_t csrCursor = 0;`
`306`	`307`	`for (size_t i = 0; i < blockSize; i++)`
`307`	`308`	`{`
`308`	`309`	`algorithmFPType minGoalVal = clustersSq[0] - x_clusters[i];`
`@@ -329,7 +330,7 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer`
`329`	`330`
`330`	`331`	`kmeansInsertCandidate(tt, minGoalVal, k * blockSizeDefault + i);`
`331`	`332`
`332`		`- *trg += minGoalVal;`
	`333`	`+ goal += minGoalVal;`
`333`	`334`
`334`	`335`	`cS0[minIdx]++;`
`335`	`336`
`@@ -339,6 +340,7 @@ Status TaskKMeansLloyd<algorithmFPType, cpu>::addNTToTaskThreadedCSR(const Numer`
`339`	`340`	`assignments[i] = (int)minIdx;`
`340`	`341`	`}`
`341`	`342`	`}`
	`343`	`+ *trg += goal;`
`342`	`344`	`});`
`343`	`345`	`return safeStat.detach();`
`344`	`346`	`}`