add single precision, add aa-vec-sl-soa kernel, updated doc

[LbmBenchmarkKernelsPublic.git] / src / BenchKernelD3Q19ListPullSplitNt2SIntrinsics.h

diff --git a/src/BenchKernelD3Q19ListPullSplitNt2SIntrinsics.h b/src/BenchKernelD3Q19ListPullSplitNt2SIntrinsics.h
index 399fa5f7b05f8ca77146b84c61fe88d70de8a323..de858ec1c0df2da90444fa81fcb9a117334c0aa6 100644 (file)
--- a/src/BenchKernelD3Q19ListPullSplitNt2SIntrinsics.h
+++ b/src/BenchKernelD3Q19ListPullSplitNt2SIntrinsics.h
@@ -41,6 +41,9 @@
 #ifdef DEBUG
                memset(tmpArray, -1, sizeof(PdfT) * nTmpArray * N_TMP);
 #endif
+               #ifdef INTEL_OPT_DIRECTIVES
+                       #pragma ivdep
+               #endif
                for (int index = 0; index < indexMax; ++index) {
 
 
@@ -69,7 +72,7 @@
                                   pdf_T  + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
                                   pdf_B  + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
 
-                       dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz) * 3.0 / 2.0;
+                       dir_indep_trm = dens - (ux * ux + uy * uy + uz * uz) * F(3.0) / F(2.0);
 
                        w_1_indep = w_1 * dir_indep_trm;
                        w_2_indep = w_2 * dir_indep_trm;