portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-14 01:35:36 +00:00
Code Releases Activity

first effort toward accelerators

Browse Source
This commit is contained in:
david clarke 2024-02-06 18:24:55 -07:00
parent 0a6e2f42c5
commit a38fb0e04a
1 changed files with 21 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
Grid/qcd/smearing/HISQSmearing.h
View File

@ -154,7 +154,7 @@ public:
for(int mu=0;mu<Nd;mu++) { for(int mu=0;mu<Nd;mu++) {
// Create the accessors // Create the accessors
autoView(U_v , Ughost , CpuRead); autoView(U_v , Ughost , AcceleratorRead);
autoView(U_fat_v , Ughost_fat , CpuWrite); autoView(U_fat_v , Ughost_fat , CpuWrite);
autoView(U_3link_v , Ughost_3link , CpuWrite); autoView(U_3link_v , Ughost_3link , CpuWrite);
autoView(U_5linkA_v, Ughost_5linkA, CpuWrite); autoView(U_5linkA_v, Ughost_5linkA, CpuWrite);
@ -171,8 +171,10 @@ public:
stencilElement SE0, SE1, SE2, SE3, SE4, SE5; stencilElement SE0, SE1, SE2, SE3, SE4, SE5;
U3matrix U0, U1, U2, U3, U4, U5, W; U3matrix U0, U1, U2, U3, U4, U5, W;
int Nsites = U_v.size();
for(int site=0;site<U_v.size();site++){ // ----------- 3-link constructs accelerator_for(site,Nsites,Simd:Nsimd(),{ // ----------- 3-link constructs
// for(int site=0;site<Nsites;site++){ // ----------- 3-link constructs
for(int nu=0;nu<Nd;nu++) { for(int nu=0;nu<Nd;nu++) {
if(nu==mu) continue; if(nu==mu) continue;
int s = stencilIndex(mu,nu); int s = stencilIndex(mu,nu);
@ -202,11 +204,18 @@ public:
// Save 3-link construct for later and add to smeared field. // Save 3-link construct for later and add to smeared field.
U_3link_v[x](nu) = W; U_3link_v[x](nu) = W;
U_fat_v[x](mu) = U_fat_v[x](mu) + lt.c_3*W;
}
}
for(int site=0;site<U_v.size();site++){ // ----------- 5-link // The index operator (x) returns the coalesced read on GPU. The view [] index returns
// a reference to the vector object. The [x](mu) returns a reference to the densely
// packed (contiguous in memory) mu-th element of the vector object. On CPU,
// coalescedRead/Write is the identity mapping assigning vector object to vector object.
// But on GPU it's non-trivial and maps scalar object to vector object and vice versa.
coalescedWrite(U_fat_v[x](mu), U_fat_v(x)(mu) + lt.c_3*W);
}
})
accelerator_for(site,Nsites,Simd:Nsimd(),{ // ----------- 5-link
// for(int site=0;site<Nsites;site++){ // ----------- 5-link
int sigmaIndex = 0; int sigmaIndex = 0;
for(int nu=0;nu<Nd;nu++) { for(int nu=0;nu<Nd;nu++) {
if(nu==mu) continue; if(nu==mu) continue;
@ -235,14 +244,14 @@ public:
U_5linkB_v[x](rho) = W; U_5linkB_v[x](rho) = W;
} }
U_fat_v[x](mu) = U_fat_v[x](mu) + lt.c_5*W; coalescedWrite(U_fat_v[x](mu), U_fat_v(x)(mu) + lt.c_5*W);
sigmaIndex++; sigmaIndex++;
} }
} }
} })
for(int site=0;site<U_v.size();site++){ // ----------- 7-link accelerator_for(site,Nsites,Simd:Nsimd(),{ // ----------- 7-link
// for(int site=0;site<Nsites;site++){ // ----------- 7-link
int sigmaIndex = 0; int sigmaIndex = 0;
for(int nu=0;nu<Nd;nu++) { for(int nu=0;nu<Nd;nu++) {
if(nu==mu) continue; if(nu==mu) continue;
@ -273,12 +282,11 @@ public:
W = U2*U1*adj(U0) + adj(U5)*U4*U3; W = U2*U1*adj(U0) + adj(U5)*U4*U3;
U_fat_v[x](mu) = U_fat_v[x](mu) + lt.c_7*W; coalescedWrite(U_fat_v[x](mu), U_fat_v(x)(mu) + lt.c_7*W);
sigmaIndex++; sigmaIndex++;
} }
} }
} })
} // end mu loop } // end mu loop