projects
/
LbmBenchmarkKernelsPublic.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
add single precision, add aa-vec-sl-soa kernel, updated doc
[LbmBenchmarkKernelsPublic.git]
/
src
/
BenchKernelD3Q19List.c
diff --git
a/src/BenchKernelD3Q19List.c
b/src/BenchKernelD3Q19List.c
index e01853aab7cf668a25f5fce990120c8d6e6c73c0..4adb858b04784a17bf9535c555614647673e2432 100644
(file)
--- a/
src/BenchKernelD3Q19List.c
+++ b/
src/BenchKernelD3Q19List.c
@@
-40,28
+40,27
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
Assert(kernelData != NULL);
Assert(cd != NULL);
Assert(kernelData != NULL);
Assert(cd != NULL);
- Assert(cd->Omega >
0.0
);
- Assert(cd->Omega <
2.0
);
+ Assert(cd->Omega >
F(0.0)
);
+ Assert(cd->Omega <
F(2.0)
);
KernelData * kd = (KernelData *)kernelData;
KernelDataList * kdl = (KernelDataList *)kernelData;
PdfT omega = cd->Omega;
PdfT omegaEven = omega;
KernelData * kd = (KernelData *)kernelData;
KernelDataList * kdl = (KernelDataList *)kernelData;
PdfT omega = cd->Omega;
PdfT omegaEven = omega;
-// PdfT omegaOdd = 8.0*((2.0-omegaEven)/(8.0-omegaEven)); //"standard" trt odd relaxation parameter
- PdfT magicParam = 1.0/12.0; // 1/4: best stability; 1/12: removes third-order advection error (best advection); 1/6: removes fourth-order diffusion error (best diffusion); 3/16: exact location of bounce back for poiseuille flow
- PdfT omegaOdd = 1.0/( 0.5 + magicParam/(1.0/omega - 0.5) );
+ PdfT magicParam = F(1.0) / F(12.0);
+ PdfT omegaOdd = F(1.0) /(F(0.5) + magicParam / (F(1.0) / omega - F(0.5)));
- PdfT evenPart =
0.0
;
- PdfT oddPart
= 0.0
;
- PdfT dir_indep_trm =
0.0
;
+ PdfT evenPart =
F(0.0)
;
+ PdfT oddPart
= F(0.0)
;
+ PdfT dir_indep_trm =
F(0.0)
;
- PdfT w_0 =
1.0 / 3.0
;
- PdfT w_1 =
1.0 / 18.0
;
- PdfT w_2 =
1.0 / 36.0
;
+ PdfT w_0 =
F(1.0) / F( 3.0)
;
+ PdfT w_1 =
F(1.0) / F(18.0)
;
+ PdfT w_2 =
F(1.0) / F(36.0)
;
- PdfT w_1_x3 = w_1 *
3.0; PdfT w_1_nine_half = w_1 * 9.0/2.0; PdfT w_1_indep = 0.0
;
- PdfT w_2_x3 = w_2 *
3.0; PdfT w_2_nine_half = w_2 * 9.0/2.0; PdfT w_2_indep = 0.0
;
+ PdfT w_1_x3 = w_1 *
F(3.0); PdfT w_1_nine_half = w_1 * F(9.0) / F(2.0); PdfT w_1_indep = F(0.0)
;
+ PdfT w_2_x3 = w_2 *
F(3.0); PdfT w_2_nine_half = w_2 * F(9.0) / F(2.0); PdfT w_2_indep = F(0.0)
;
PdfT ux, uy, uz, ui;
PdfT dens;
PdfT ux, uy, uz, ui;
PdfT dens;
@@
-112,6
+111,9
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
evenPart, oddPart, w_1_indep, w_2_indep)
#endif
pdf_B, pdf_BN, pdf_BE, pdf_BS, pdf_BW, \
evenPart, oddPart, w_1_indep, w_2_indep)
#endif
+ #ifdef INTEL_OPT_DIRECTIVES
+ #pragma ivdep
+ #endif
for (int index = 0; index < nFluid; ++index) {
#define I(index, dir) P_INDEX_3((nCells), (index), (dir))
for (int index = 0; index < nFluid; ++index) {
#define I(index, dir) P_INDEX_3((nCells), (index), (dir))
@@
-171,7
+173,7
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
pdf_T + pdf_TN + pdf_TE + pdf_TS + pdf_TW +
pdf_B + pdf_BN + pdf_BE + pdf_BS + pdf_BW;
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)
;
#ifdef PROP_MODEL_PUSH
#ifdef PROP_MODEL_PUSH
@@
-184,20
+186,20
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
w_1_indep = w_1*dir_indep_trm;
ui = uy;
w_1_indep = w_1*dir_indep_trm;
ui = uy;
- evenPart = omegaEven*(
0.5
*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_N - pdf_S) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_N - pdf_S) - ui*w_1_x3 );
dst[adjList[adjListIndex + D3Q19_N]] = pdf_N - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_S]] = pdf_S - evenPart + oddPart;
ui = ux;
dst[adjList[adjListIndex + D3Q19_N]] = pdf_N - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_S]] = pdf_S - evenPart + oddPart;
ui = ux;
- evenPart = omegaEven*(
0.5
*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_E - pdf_W) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_E - pdf_W) - ui*w_1_x3 );
dst[adjList[adjListIndex + D3Q19_E]] = pdf_E - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_W]] = pdf_W - evenPart + oddPart;
ui = uz;
dst[adjList[adjListIndex + D3Q19_E]] = pdf_E - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_W]] = pdf_W - evenPart + oddPart;
ui = uz;
- evenPart = omegaEven*(
0.5
*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_T - pdf_B) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_T - pdf_B) - ui*w_1_x3 );
dst[adjList[adjListIndex + D3Q19_T]] = pdf_T - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_B]] = pdf_B - evenPart + oddPart;
dst[adjList[adjListIndex + D3Q19_T]] = pdf_T - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_B]] = pdf_B - evenPart + oddPart;
@@
-205,38
+207,38
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
w_2_indep = w_2*dir_indep_trm;
ui = -ux + uy;
w_2_indep = w_2*dir_indep_trm;
ui = -ux + uy;
- evenPart = omegaEven*(
0.5
*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_NW - pdf_SE) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_NW - pdf_SE) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_NW]] = pdf_NW - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_SE]] = pdf_SE - evenPart + oddPart;
ui = ux + uy;
dst[adjList[adjListIndex + D3Q19_NW]] = pdf_NW - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_SE]] = pdf_SE - evenPart + oddPart;
ui = ux + uy;
- evenPart = omegaEven*(
0.5
*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_NE - pdf_SW) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_NE - pdf_SW) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_NE]] = pdf_NE - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_SW]] = pdf_SW - evenPart + oddPart;
ui = -ux + uz;
dst[adjList[adjListIndex + D3Q19_NE]] = pdf_NE - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_SW]] = pdf_SW - evenPart + oddPart;
ui = -ux + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TW - pdf_BE) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TW - pdf_BE) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_TW]] = pdf_TW - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BE]] = pdf_BE - evenPart + oddPart;
ui = ux + uz;
dst[adjList[adjListIndex + D3Q19_TW]] = pdf_TW - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BE]] = pdf_BE - evenPart + oddPart;
ui = ux + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TE - pdf_BW) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TE - pdf_BW) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_TE]] = pdf_TE - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BW]] = pdf_BW - evenPart + oddPart;
ui = -uy + uz;
dst[adjList[adjListIndex + D3Q19_TE]] = pdf_TE - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BW]] = pdf_BW - evenPart + oddPart;
ui = -uy + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TS - pdf_BN) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TS - pdf_BN) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_TS]] = pdf_TS - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BN]] = pdf_BN - evenPart + oddPart;
ui = uy + uz;
dst[adjList[adjListIndex + D3Q19_TS]] = pdf_TS - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BN]] = pdf_BN - evenPart + oddPart;
ui = uy + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TN - pdf_BS) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TN - pdf_BS) - ui*w_2_x3 );
dst[adjList[adjListIndex + D3Q19_TN]] = pdf_TN - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BS]] = pdf_BS - evenPart + oddPart;
dst[adjList[adjListIndex + D3Q19_TN]] = pdf_TN - evenPart - oddPart;
dst[adjList[adjListIndex + D3Q19_BS]] = pdf_BS - evenPart + oddPart;
@@
-249,20
+251,20
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
w_1_indep = w_1*dir_indep_trm;
ui = uy;
w_1_indep = w_1*dir_indep_trm;
ui = uy;
- evenPart = omegaEven*(
0.5
*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_N - pdf_S) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_N + pdf_S) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_N - pdf_S) - ui*w_1_x3 );
dst[I(index, D3Q19_N )] = pdf_N - evenPart - oddPart;
dst[I(index, D3Q19_S )] = pdf_S - evenPart + oddPart;
ui = ux;
dst[I(index, D3Q19_N )] = pdf_N - evenPart - oddPart;
dst[I(index, D3Q19_S )] = pdf_S - evenPart + oddPart;
ui = ux;
- evenPart = omegaEven*(
0.5
*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_E - pdf_W) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_E + pdf_W) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_E - pdf_W) - ui*w_1_x3 );
dst[I(index, D3Q19_E )] = pdf_E - evenPart - oddPart;
dst[I(index, D3Q19_W )] = pdf_W - evenPart + oddPart;
ui = uz;
dst[I(index, D3Q19_E )] = pdf_E - evenPart - oddPart;
dst[I(index, D3Q19_W )] = pdf_W - evenPart + oddPart;
ui = uz;
- evenPart = omegaEven*(
0.5
*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_T - pdf_B) - ui*w_1_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_T + pdf_B) - ui*ui*w_1_nine_half - w_1_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_T - pdf_B) - ui*w_1_x3 );
dst[I(index, D3Q19_T )] = pdf_T - evenPart - oddPart;
dst[I(index, D3Q19_B )] = pdf_B - evenPart + oddPart;
dst[I(index, D3Q19_T )] = pdf_T - evenPart - oddPart;
dst[I(index, D3Q19_B )] = pdf_B - evenPart + oddPart;
@@
-270,38
+272,38
@@
void FNAME(D3Q19ListKernel)(LatticeDesc * ld, KernelData * kernelData, CaseData
w_2_indep = w_2*dir_indep_trm;
ui = -ux + uy;
w_2_indep = w_2*dir_indep_trm;
ui = -ux + uy;
- evenPart = omegaEven*(
0.5
*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_NW - pdf_SE) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_NW + pdf_SE) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_NW - pdf_SE) - ui*w_2_x3 );
dst[I(index, D3Q19_NW)] = pdf_NW - evenPart - oddPart;
dst[I(index, D3Q19_SE)] = pdf_SE - evenPart + oddPart;
ui = ux + uy;
dst[I(index, D3Q19_NW)] = pdf_NW - evenPart - oddPart;
dst[I(index, D3Q19_SE)] = pdf_SE - evenPart + oddPart;
ui = ux + uy;
- evenPart = omegaEven*(
0.5
*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_NE - pdf_SW) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_NE + pdf_SW) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_NE - pdf_SW) - ui*w_2_x3 );
dst[I(index, D3Q19_NE)] = pdf_NE - evenPart - oddPart;
dst[I(index, D3Q19_SW)] = pdf_SW - evenPart + oddPart;
ui = -ux + uz;
dst[I(index, D3Q19_NE)] = pdf_NE - evenPart - oddPart;
dst[I(index, D3Q19_SW)] = pdf_SW - evenPart + oddPart;
ui = -ux + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TW - pdf_BE) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TW + pdf_BE) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TW - pdf_BE) - ui*w_2_x3 );
dst[I(index, D3Q19_TW)] = pdf_TW - evenPart - oddPart;
dst[I(index, D3Q19_BE)] = pdf_BE - evenPart + oddPart;
ui = ux + uz;
dst[I(index, D3Q19_TW)] = pdf_TW - evenPart - oddPart;
dst[I(index, D3Q19_BE)] = pdf_BE - evenPart + oddPart;
ui = ux + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TE - pdf_BW) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TE + pdf_BW) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TE - pdf_BW) - ui*w_2_x3 );
dst[I(index, D3Q19_TE)] = pdf_TE - evenPart - oddPart;
dst[I(index, D3Q19_BW)] = pdf_BW - evenPart + oddPart;
ui = -uy + uz;
dst[I(index, D3Q19_TE)] = pdf_TE - evenPart - oddPart;
dst[I(index, D3Q19_BW)] = pdf_BW - evenPart + oddPart;
ui = -uy + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TS - pdf_BN) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TS + pdf_BN) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TS - pdf_BN) - ui*w_2_x3 );
dst[I(index, D3Q19_TS)] = pdf_TS - evenPart - oddPart;
dst[I(index, D3Q19_BN)] = pdf_BN - evenPart + oddPart;
ui = uy + uz;
dst[I(index, D3Q19_TS)] = pdf_TS - evenPart - oddPart;
dst[I(index, D3Q19_BN)] = pdf_BN - evenPart + oddPart;
ui = uy + uz;
- evenPart = omegaEven*(
0.5
*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
- oddPart = omegaOdd*(
0.5
*(pdf_TN - pdf_BS) - ui*w_2_x3 );
+ evenPart = omegaEven*(
F(0.5)
*(pdf_TN + pdf_BS) - ui*ui*w_2_nine_half - w_2_indep );
+ oddPart = omegaOdd*(
F(0.5)
*(pdf_TN - pdf_BS) - ui*w_2_x3 );
dst[I(index, D3Q19_TN)] = pdf_TN - evenPart - oddPart;
dst[I(index, D3Q19_BS)] = pdf_BS - evenPart + oddPart;
dst[I(index, D3Q19_TN)] = pdf_TN - evenPart - oddPart;
dst[I(index, D3Q19_BS)] = pdf_BS - evenPart + oddPart;
This page took
0.054154 seconds
and
5
git commands to generate.