Skip to content

Commit

Permalink
Fix a race condition in splitVertices
Browse files Browse the repository at this point in the history 
Add

    alpaka::syncBlockThreads(acc);

at the end of the loop on the vertices to ensure that all threads are
properly synchronised before resetting the shared memory.

Clean up the kernel to use the SoA accessors and the cms::alpakatools
utilities.
  • Loading branch information
@fwyzard
fwyzard committed Aug 6, 2024
1 parent f6e0fc4 commit e746fdf
Showing 1 changed file with 33 additions and 48 deletions.
81 changes: 33 additions & 48 deletions RecoTracker/PixelVertexFinding/plugins/alpaka/splitVertices.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,30 +19,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder {
using WsSoAView = ::vertexFinder::PixelVertexWorkSpaceSoAView;
template <typename TAcc>
ALPAKA_FN_ACC ALPAKA_FN_INLINE __attribute__((always_inline)) void splitVertices(const TAcc& acc,
VtxSoAView& pdata,
WsSoAView& pws,
VtxSoAView& data,
WsSoAView& ws,
float maxChi2) {
constexpr bool verbose = false; // in principle the compiler should optmize out if false
const uint32_t threadIdxLocal(alpaka::getIdx<alpaka::Block, alpaka::Threads>(acc)[0u]);

auto& __restrict__ data = pdata;
auto& __restrict__ ws = pws;
auto nt = ws.ntrks();
float const* __restrict__ zt = ws.zt();
float const* __restrict__ ezt2 = ws.ezt2();
float* __restrict__ zv = data.zv();
float* __restrict__ wv = data.wv();
float const* __restrict__ chi2 = data.chi2();
uint32_t& nvFinal = data.nvFinal();

int32_t const* __restrict__ nn = data.ndof();
int32_t* __restrict__ iv = ws.iv();

ALPAKA_ASSERT_ACC(zt);
ALPAKA_ASSERT_ACC(wv);
ALPAKA_ASSERT_ACC(chi2);
ALPAKA_ASSERT_ACC(nn);

constexpr uint32_t MAXTK = 512;

auto& it = alpaka::declareSharedVar<uint32_t[MAXTK], __COUNTER__>(acc); // track index
Expand All @@ -51,32 +31,33 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder {
auto& ww = alpaka::declareSharedVar<float[MAXTK], __COUNTER__>(acc); // z weight
auto& nq = alpaka::declareSharedVar<uint32_t, __COUNTER__>(acc); // number of track for this vertex

const uint32_t blockIdx(alpaka::getIdx<alpaka::Grid, alpaka::Blocks>(acc)[0u]);
const uint32_t gridDimension(alpaka::getWorkDiv<alpaka::Grid, alpaka::Blocks>(acc)[0u]);

// one vertex per block
for (auto kv = blockIdx; kv < nvFinal; kv += gridDimension) {
if (nn[kv] < 4)
for (auto kv : cms::alpakatools::independent_groups(acc, data.nvFinal())) {
int32_t ndof = data[kv].ndof();
if (ndof < 4)
continue;
if (chi2[kv] < maxChi2 * float(nn[kv]))
if (data[kv].chi2() < maxChi2 * float(ndof))
continue;

ALPAKA_ASSERT_ACC(nn[kv] < int32_t(MAXTK));
ALPAKA_ASSERT_ACC(ndof < int32_t(MAXTK));

if ((uint32_t)nn[kv] >= MAXTK)
if ((uint32_t)ndof >= MAXTK)
continue; // too bad FIXME

nq = 0u;
if (cms::alpakatools::once_per_block(acc)) {
// reset the number of tracks for the current vertex
nq = 0u;
}
alpaka::syncBlockThreads(acc);

// copy to local
for (auto k : cms::alpakatools::independent_group_elements(acc, nt)) {
if (iv[k] == int(kv)) {
auto old = alpaka::atomicInc(acc, &nq, MAXTK, alpaka::hierarchy::Threads{});
zz[old] = zt[k] - zv[kv];
newV[old] = zz[old] < 0 ? 0 : 1;
ww[old] = 1.f / ezt2[k];
it[old] = k;
// cache the data of the tracks associated to the current vertex into shared memory
for (auto k : cms::alpakatools::independent_group_elements(acc, ws.ntrks())) {
if (ws[k].iv() == int(kv)) {
auto index = alpaka::atomicInc(acc, &nq, MAXTK, alpaka::hierarchy::Threads{});
it[index] = k;
zz[index] = ws[k].zt() - data[kv].zv();
newV[index] = zz[index] < 0 ? 0 : 1;
ww[index] = 1.f / ws[k].ezt2();
}
}

Expand All @@ -85,14 +66,14 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder {
auto& wnew = alpaka::declareSharedVar<float[2], __COUNTER__>(acc);
alpaka::syncBlockThreads(acc);

ALPAKA_ASSERT_ACC(int(nq) == nn[kv] + 1);
ALPAKA_ASSERT_ACC(int(nq) == ndof + 1);

int maxiter = 20;
// kt-min....
bool more = true;
while (alpaka::syncBlockThreadsPredicate<alpaka::BlockOr>(acc, more)) {
more = false;
if (0 == threadIdxLocal) {
if (cms::alpakatools::once_per_block(acc)) {
znew[0] = 0;
znew[1] = 0;
wnew[0] = 0;
Expand All @@ -107,7 +88,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder {
}
alpaka::syncBlockThreads(acc);

if (0 == threadIdxLocal) {
if (cms::alpakatools::once_per_block(acc)) {
znew[0] /= wnew[0];
znew[1] /= wnew[1];
}
Expand All @@ -134,30 +115,34 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder {

auto chi2Dist = dist2 / (1.f / wnew[0] + 1.f / wnew[1]);

if (verbose && 0 == threadIdxLocal)
printf("inter %d %f %f\n", 20 - maxiter, chi2Dist, dist2 * wv[kv]);
if constexpr (verbose) {
if (cms::alpakatools::once_per_block(acc))
printf("inter %d %f %f\n", 20 - maxiter, chi2Dist, dist2 * data[kv].wv());
}

if (chi2Dist < 4)
continue;

// get a new global vertex
auto& igv = alpaka::declareSharedVar<uint32_t, __COUNTER__>(acc);
if (0 == threadIdxLocal)
if (cms::alpakatools::once_per_block(acc))
igv = alpaka::atomicAdd(acc, &ws.nvIntermediate(), 1u, alpaka::hierarchy::Blocks{});
alpaka::syncBlockThreads(acc);
for (auto k : cms::alpakatools::uniform_elements(acc, nq)) {
if (1 == newV[k])
iv[it[k]] = igv;
ws[it[k]].iv() = igv;
}

// synchronise the threads before starting the next iteration of the loop over the vertices and resetting the shared memory
alpaka::syncBlockThreads(acc);
} // loop on vertices
}

class SplitVerticesKernel {
public:
template <typename TAcc>
ALPAKA_FN_ACC void operator()(const TAcc& acc, VtxSoAView pdata, WsSoAView pws, float maxChi2) const {
splitVertices(acc, pdata, pws, maxChi2);
ALPAKA_FN_ACC void operator()(const TAcc& acc, VtxSoAView data, WsSoAView ws, float maxChi2) const {
splitVertices(acc, data, ws, maxChi2);
}
};

Expand Down

0 comments on commit e746fdf

Please sign in to comment.