AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

rocm-smi and KokkosTools support (#23)

Browse Source 
* renamed omnitrace_thread_data to thread_data

* initial implementation

* Numerous fixes and updates

- Updated timemory submodule
- Updated perfetto submodule (pulls in fixes for TRACE_EVENT)
- pthread_gotcha only after omnitrace_init_tooling
- omnitrace banner
- config settings for rocm-smi freq and devices
- critical_trace::get_entries
- OMNITRACE_BASIC_PRINT
- rocm_smi perfetto category
- redirect roctracer warnings for ROCm 4.5.0
- property specializations for rocm-smi components
- units fixes data_tracker types
- roctracer entries for pthread_create and start_thread
- omnitrace-avail defaults to settings, not components
- settings have conforming names
- settings warn about duplicates
- ptl named threads
- decreased max freq for sampler SIGALRM
- rocm-smi names thread
- rocm-smi avoids call to hipGetDeviceCount
- name roctracer activity callback threads
- fixed binary rewrite test output names

* Update lulesh example

- supports non-UVM GPU

* Lulesh tweaks + formatting

* KokkosP + Mode + Roctracer sampling deadlock fix

- kokkosp support
- omnitrace_init_library
- config::print_settings()
- config::get_mode()
- omnitrace::Mode
- omnitrace-avail improvements (removes settings)
- handle get_verbose() < 0
- disable dyninst InstrStackFrames by default
- handle perf_event_paranoid > 1 by disabling PAPI
- SIGALRM max freq to 5.0
- Name threads
- rocm-smi handles get_use_perfetto() and get_use_timemory()
- HSA_ENABLE_INTERRUPT=0 when roctracer + sampling (fixes deadlock)

* Tests, API renaming, roctracer

- disable renaming of thread 0
- verbprintf_bare
- enable dyninst merge tramp
- tweaked some omnitrace exe verbose levels
- reworked roctracer::setup and roctracer::shutdown
- rocm_smi::data::poll checks get_state()
- omnitrace_trace_finalize -> omnitrace_finalize
- omnitrace_trace_init -> omnitrace_init
- omnitrace_trace_set_env -> omnitrace_set_env
- omnitrace_trace_set_mpi -> omnitrace_set_mpi
- sampling mode does not disable timemory
- disable roctracer before shutting down rocm-smi
- lulesh tests w/ and w/o kokkosp
- lulesh tests for perfetto only
    - with --dynamic-callsites --traps --allow-overlapping
- lulesh tests for timemory only
    - with --stdlib --dynamic-callsites --traps --allow-overlapping

* Update timemory submodule

- fix for TIMEMORY_PROPERTY_SPECIALIZATION

* get_verbose() handling + timemory submodule update

- Findroctracer.cmake uses find_package(hsakmt)

* Stability fixes + rework roctracer + perfetto

- reworked roctracer start up
- critical_trace perfetto basic values
- perfetto sampling category
- sampler checks signals
- peak_rss in sampling
- pthread_gotcha::shutdown()
- rocm_smi::device_count()
- HSA_TOOLS_LIB is set
- HSA_ENABLE_INTERRUPT in omnitrace exe
- omnitrace exe verbosity level changes
- Avoid instrumenting Impl ns in Kokkos
- gpu::device_count prefers rocm_smi instead of hip
- ptl blocks signals
- fixed pthread_gotcha roctracer_data values
- removed runtime-instrument-sampling tests
- timemory submodule update

* cmake formatting

* timemory + roctracer updates

- fix timemory issue with papi_common
- fix timemory issue with units
- define roctracer::is_setup()

* Miscellaneous tweaks

- Disable sampling during runtime instrument
- Fixed warnings about dynamic callsites
- Fixed backtrace output when timemory disabled
- Test tweaks

* cmake-format

* omnitrace_target_compile_definitions

* timemory submodule update

* config, omnitrace, State, mpi_gotcha updates

- use OMNITRACE_THROW instead of direct throw
- is_attached()
- is_binary_rewrite()
- get_is_continuous_integration()
- get_debug_init()
- get_debug_finalize()
- max_thread_bookmarks default to 1
- State::Init
- app_thread oneTimeCode
- runtime instrumentation uses waitpid
- fixed init_names
- include main in MPI runs
- fixed sampling setup when disabled
- reworked mpi_gotcha
- disabled critical trace in transpose test

* cmake-format

* handle rocm_smi::device_count() exception

* CI timeouts

* Re-enable runtime-instrument + sampling

[ROCm/rocprofiler-systems commit: 39f17ae8b8]
This commit is contained in:
Jonathan R. Madsen
2022-02-08 17:42:17 -06:00
committed by GitHub
parent b4a82711d1
commit 4ae26e2d08
59 changed files with 5116 additions and 2592 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-systems/examples/lulesh/CMakeLists.txt
+13 -2
View File
@@ -4,7 +4,16 @@ project(lulesh LANGUAGES C CXX)
list(INSERT CMAKE_MODULE_PATH 0 ${PROJECT_SOURCE_DIR}/cmake/Modules)
add_subdirectory(external)
option(LULESH_BUILD_KOKKOS "Build Kokkos from submodule" ON)
if(LULESH_BUILD_KOKKOS)
add_subdirectory(external)
if(LULESH_USE_CUDA OR LULESH_USE_HIP)
kokkos_compilation(PROJECT)
endif()
else()
find_package(Kokkos REQUIRED COMPONENTS separable_compilation)
kokkos_compilation(PROJECT)
endif()
set(CMAKE_CXX_EXTENSIONS OFF)
@@ -14,7 +23,9 @@ if("${CMAKE_BUILD_TYPE}" STREQUAL "")
CACHE STRING "CMake build type" FORCE)
endif()
if(DEFINED OMNITRACE_USE_MPI)
if(DEFINED OMNITRACE_USE_MPI
AND NOT LULESH_USE_CUDA
AND NOT LULESH_USE_HIP)
option(LULESH_USE_MPI "Enable MPI" ${OMNITRACE_USE_MPI})
else()
option(LULESH_USE_MPI "Enable MPI" OFF)
projects/rocprofiler-systems/examples/lulesh/external/CMakeLists.txt
Vendored
+16 -11
View File
@@ -1,22 +1,27 @@
# include(Utilities)
option(LULESH_USE_CUDA "Enable Kokkos CUDA backend for lulesh" OFF)
option(LULESH_USE_HIP "Enable Kokkos HIP backend for lulesh" OFF)
set(Kokkos_ENABLE_SERIAL
ON
CACHE BOOL "Enable Serial")
set(Kokkos_ENABLE_OPENMP
ON
CACHE BOOL "Enable OpenMP")
if(USE_CUDA)
if(LULESH_USE_CUDA)
set(Kokkos_ENABLE_CUDA
ON
CACHE BOOL "Enable CUDA")
set(Kokkos_ENABLE_CUDA_UVM
ON
CACHE BOOL "Enable CUDA UVM")
CACHE BOOL "Enable CUDA" FORCE)
set(Kokkos_ENABLE_CUDA_LAMBDA
ON
CACHE BOOL "Enable CUDA UVM")
set(Kokkos_ENABLE_CUDA_CONSTEXPR
CACHE BOOL "Enable CUDA lambda support" FORCE)
elseif(LULESH_USE_HIP)
set(Kokkos_ENABLE_HIP
ON
CACHE BOOL "Enable CUDA UVM")
CACHE BOOL "Enable HIP" FORCE)
else()
set(Kokkos_ENABLE_OPENMP
ON
CACHE BOOL "Enable OpenMP")
endif()
checkout_git_submodule(
projects/rocprofiler-systems/examples/lulesh/external/kokkos
Vendored
+1 -1
Submodule projects/rocprofiler-systems/examples/lulesh/external/kokkos updated: 56468253ef...698a67731a
projects/rocprofiler-systems/examples/lulesh/lulesh-comm.cc
+7 -12
View File
@@ -9,8 +9,8 @@
/* Comm Routines */
# define ALLOW_UNPACKED_PLANE false
# define ALLOW_UNPACKED_ROW false
# define ALLOW_UNPACKED_COL false
# define ALLOW_UNPACKED_ROW false
# define ALLOW_UNPACKED_COL false
/*
There are coherence issues for packing and unpacking message
@@ -59,8 +59,7 @@ void
CommRecv(Domain& domain, int msgType, Index_t xferFields, Index_t dx, Index_t dy,
Index_t dz, bool doRecv, bool planeOnly)
{
if(domain.numRanks() == 1)
return;
if(domain.numRanks() == 1) return;
/* post recieve buffers for all incoming messages */
int myRank;
@@ -366,8 +365,7 @@ void
CommSend(Domain& domain, int msgType, Index_t xferFields, Domain_member* fieldData,
Index_t dx, Index_t dy, Index_t dz, bool doSend, bool planeOnly)
{
if(domain.numRanks() == 1)
return;
if(domain.numRanks() == 1) return;
/* post recieve buffers for all incoming messages */
int myRank;
@@ -926,8 +924,7 @@ CommSend(Domain& domain, int msgType, Index_t xferFields, Domain_member* fieldDa
void
CommSBN(Domain& domain, int xferFields, Domain_member* fieldData)
{
if(domain.numRanks() == 1)
return;
if(domain.numRanks() == 1) return;
/* summation order should be from smallest value to largest */
/* or we could try out kahan summation! */
@@ -1405,8 +1402,7 @@ CommSBN(Domain& domain, int xferFields, Domain_member* fieldData)
void
CommSyncPosVel(Domain& domain)
{
if(domain.numRanks() == 1)
return;
if(domain.numRanks() == 1) return;
int myRank;
bool doRecv = false;
@@ -1893,8 +1889,7 @@ CommSyncPosVel(Domain& domain)
void
CommMonoQ(Domain& domain)
{
if(domain.numRanks() == 1)
return;
if(domain.numRanks() == 1) return;
int myRank;
Index_t xferFields = 3; /* delv_xi, delv_eta, delv_zeta */
projects/rocprofiler-systems/examples/lulesh/lulesh-init.cc
+196 -151
View File
@@ -2,9 +2,6 @@
#if USE_MPI
# include <mpi.h>
#endif
#if _OPENMP
# include <omp.h>
#endif
#include "lulesh.h"
#include <cstdlib>
#include <limits.h>
@@ -13,12 +10,12 @@
#include <string.h>
static KOKKOS_INLINE_FUNCTION Real_t
CalcElemVolume(const Real_t x0, const Real_t x1, const Real_t x2, const Real_t x3,
const Real_t x4, const Real_t x5, const Real_t x6, const Real_t x7,
const Real_t y0, const Real_t y1, const Real_t y2, const Real_t y3,
const Real_t y4, const Real_t y5, const Real_t y6, const Real_t y7,
const Real_t z0, const Real_t z1, const Real_t z2, const Real_t z3,
const Real_t z4, const Real_t z5, const Real_t z6, const Real_t z7)
CalcElemVolume(const Real_t x0, const Real_t x1, const Real_t x2, const Real_t x3,
const Real_t x4, const Real_t x5, const Real_t x6, const Real_t x7,
const Real_t y0, const Real_t y1, const Real_t y2, const Real_t y3,
const Real_t y4, const Real_t y5, const Real_t y6, const Real_t y7,
const Real_t z0, const Real_t z1, const Real_t z2, const Real_t z3,
const Real_t z4, const Real_t z5, const Real_t z6, const Real_t z7)
{
Real_t twelveth = Real_t(1.0) / Real_t(12.0);
@@ -122,15 +119,15 @@ Domain::Domain(Int_t numRanks, Index_t colLoc, Index_t rowLoc, Index_t planeLoc,
, m_dvovmax(Real_t(0.1))
, m_refdens(Real_t(1.0))
,
//
// set pointers to (potentially) "new'd" arrays to null to
// simplify deallocation.
//
m_regNumList(0)
, m_nodeElemStart(0)
, m_nodeElemCornerList(0)
, m_regElemSize(0)
, m_regElemlist(0)
//
// set pointers to (potentially) "new'd" arrays to null to
// simplify deallocation.
//
m_regNumList(0)
// m_nodeElemStart(0),
// m_nodeElemCornerList(0),
// m_regElemSize(0),
// m_regElemlist(0)
#if USE_MPI
, commDataSend(0)
, commDataRecv(0)
@@ -171,48 +168,27 @@ Domain::Domain(Int_t numRanks, Index_t colLoc, Index_t rowLoc, Index_t planeLoc,
SetupCommBuffers(edgeNodes);
// Basic Field Initialization
for(Index_t i = 0; i < numElem(); ++i)
{
e(i) = Real_t(0.0);
p(i) = Real_t(0.0);
q(i) = Real_t(0.0);
ss(i) = Real_t(0.0);
}
Kokkos::deep_copy(m_e, 0.0);
Kokkos::deep_copy(m_p, 0.0);
Kokkos::deep_copy(m_q, 0.0);
Kokkos::deep_copy(m_ss, 0.0);
// Note - v initializes to 1.0, not 0.0!
for(Index_t i = 0; i < numElem(); ++i)
{
v(i) = Real_t(1.0);
}
Kokkos::deep_copy(m_v, 1.0);
for(Index_t i = 0; i < numNode(); ++i)
{
xd(i) = Real_t(0.0);
yd(i) = Real_t(0.0);
zd(i) = Real_t(0.0);
}
Kokkos::deep_copy(m_xd, 0.0);
Kokkos::deep_copy(m_yd, 0.0);
Kokkos::deep_copy(m_zd, 0.0);
for(Index_t i = 0; i < numNode(); ++i)
{
xdd(i) = Real_t(0.0);
ydd(i) = Real_t(0.0);
zdd(i) = Real_t(0.0);
}
Kokkos::deep_copy(m_xdd, 0.0);
Kokkos::deep_copy(m_ydd, 0.0);
Kokkos::deep_copy(m_zdd, 0.0);
for(Index_t i = 0; i < numNode(); ++i)
{
nodalMass(i) = Real_t(0.0);
}
Kokkos::deep_copy(m_nodalMass, 0.0);
BuildMesh(nx, edgeNodes, edgeElems);
#if _OPENMP
SetupThreadSupportStructures();
#else
// These arrays are not used if we're not threaded
m_nodeElemStart = NULL;
m_nodeElemCornerList = NULL;
#endif
// Setup region index sets. For now, these are constant sized
// throughout the run, but could be changed every cycle to
@@ -247,30 +223,46 @@ Domain::Domain(Int_t numRanks, Index_t colLoc, Index_t rowLoc, Index_t planeLoc,
time() = Real_t(0.);
cycle() = Int_t(0);
// With C++17 requirement we could just run this on the device
// without creating temporary host copies
auto h_nodelist = Kokkos::create_mirror_view(m_nodelist);
auto h_x = Kokkos::create_mirror_view(m_x);
auto h_y = Kokkos::create_mirror_view(m_y);
auto h_z = Kokkos::create_mirror_view(m_z);
auto h_volo = Kokkos::create_mirror_view(m_volo);
auto h_elemMass = Kokkos::create_mirror_view(m_elemMass);
auto h_nodalMass = Kokkos::create_mirror_view(m_nodalMass);
Kokkos::deep_copy(h_nodelist, m_nodelist);
Kokkos::deep_copy(h_x, m_x);
Kokkos::deep_copy(h_y, m_y);
Kokkos::deep_copy(h_z, m_z);
// initialize field data
for(Index_t i = 0; i < numElem(); ++i)
{
Real_t x_local[8], y_local[8], z_local[8];
Index_t* elemToNode = nodelist(i);
Real_t x_local[8], y_local[8], z_local[8];
for(Index_t lnode = 0; lnode < 8; ++lnode)
{
Index_t gnode = elemToNode[lnode];
x_local[lnode] = x(gnode);
y_local[lnode] = y(gnode);
z_local[lnode] = z(gnode);
Index_t gnode = h_nodelist(i, lnode);
x_local[lnode] = h_x(gnode);
y_local[lnode] = h_y(gnode);
z_local[lnode] = h_z(gnode);
}
// volume calculations
Real_t volume = CalcElemVolume(x_local, y_local, z_local);
volo(i) = volume;
elemMass(i) = volume;
h_volo(i) = volume;
h_elemMass(i) = volume;
for(Index_t j = 0; j < 8; ++j)
{
Index_t idx = elemToNode[j];
nodalMass(idx) += volume / Real_t(8.0);
Index_t idx = h_nodelist(i, j);
h_nodalMass(idx) += volume / Real_t(8.0);
}
}
Kokkos::deep_copy(m_volo, h_volo);
Kokkos::deep_copy(m_elemMass, h_elemMass);
Kokkos::deep_copy(m_nodalMass, h_nodalMass);
// deposit initial energy
// An energy of 3.948746e+7 is correct for a problem with
// 45 zones along a side - we need to scale it
@@ -281,10 +273,11 @@ Domain::Domain(Int_t numRanks, Index_t colLoc, Index_t rowLoc, Index_t planeLoc,
{
// Dump into the first zone (which we know is in the corner)
// of the domain that sits at the origin
e(0) = einit;
Kokkos::deep_copy(Kokkos::subview(m_e, 0), einit);
// e(0) = einit;
}
// set initial deltatime base on analytic CFL calculation
deltatime() = (Real_t(.5) * cbrt(volo(0))) / sqrt(Real_t(2.0) * einit);
deltatime() = (Real_t(.5) * cbrt(h_volo(0))) / sqrt(Real_t(2.0) * einit);
} // End constructor
@@ -315,6 +308,10 @@ Domain::BuildMesh(Int_t nx, Int_t edgeNodes, Int_t edgeElems)
{
Index_t meshEdgeElems = m_tp * nx;
auto h_x = Kokkos::create_mirror_view(m_x);
auto h_y = Kokkos::create_mirror_view(m_y);
auto h_z = Kokkos::create_mirror_view(m_z);
// initialize nodal coordinates
Index_t nidx = 0;
Real_t tz = Real_t(1.125) * Real_t(m_planeLoc * nx) / Real_t(meshEdgeElems);
@@ -326,9 +323,9 @@ Domain::BuildMesh(Int_t nx, Int_t edgeNodes, Int_t edgeElems)
Real_t tx = Real_t(1.125) * Real_t(m_colLoc * nx) / Real_t(meshEdgeElems);
for(Index_t col = 0; col < edgeNodes; ++col)
{
x(nidx) = tx;
y(nidx) = ty;
z(nidx) = tz;
h_x(nidx) = tx;
h_y(nidx) = ty;
h_z(nidx) = tz;
++nidx;
// tx += ds ; // may accumulate roundoff...
tx = Real_t(1.125) * Real_t(m_colLoc * nx + col + 1) /
@@ -341,6 +338,11 @@ Domain::BuildMesh(Int_t nx, Int_t edgeNodes, Int_t edgeElems)
tz = Real_t(1.125) * Real_t(m_planeLoc * nx + plane + 1) / Real_t(meshEdgeElems);
}
Kokkos::deep_copy(m_x, h_x);
Kokkos::deep_copy(m_y, h_y);
Kokkos::deep_copy(m_z, h_z);
auto h_nodelist = Kokkos::create_mirror_view(m_nodelist);
// embed hexehedral elements in nodal point lattice
Index_t zidx = 0;
nidx = 0;
@@ -350,15 +352,14 @@ Domain::BuildMesh(Int_t nx, Int_t edgeNodes, Int_t edgeElems)
{
for(Index_t col = 0; col < edgeElems; ++col)
{
Index_t* localNode = nodelist(zidx);
localNode[0] = nidx;
localNode[1] = nidx + 1;
localNode[2] = nidx + edgeNodes + 1;
localNode[3] = nidx + edgeNodes;
localNode[4] = nidx + edgeNodes * edgeNodes;
localNode[5] = nidx + edgeNodes * edgeNodes + 1;
localNode[6] = nidx + edgeNodes * edgeNodes + edgeNodes + 1;
localNode[7] = nidx + edgeNodes * edgeNodes + edgeNodes;
h_nodelist(zidx, 0) = nidx;
h_nodelist(zidx, 1) = nidx + 1;
h_nodelist(zidx, 2) = nidx + edgeNodes + 1;
h_nodelist(zidx, 3) = nidx + edgeNodes;
h_nodelist(zidx, 4) = nidx + edgeNodes * edgeNodes;
h_nodelist(zidx, 5) = nidx + edgeNodes * edgeNodes + 1;
h_nodelist(zidx, 6) = nidx + edgeNodes * edgeNodes + edgeNodes + 1;
h_nodelist(zidx, 7) = nidx + edgeNodes * edgeNodes + edgeNodes;
++zidx;
++nidx;
}
@@ -366,6 +367,7 @@ Domain::BuildMesh(Int_t nx, Int_t edgeNodes, Int_t edgeElems)
}
nidx += edgeNodes;
}
Kokkos::deep_copy(m_nodelist, h_nodelist);
}
////////////////////////////////////////////////////////////////////////////////
@@ -373,32 +375,31 @@ void
Domain::SetupThreadSupportStructures()
{
// set up node-centered indexing of elements
Index_t* nodeElemCount = Allocate<Index_t>(numNode());
for(Index_t i = 0; i < numNode(); ++i)
{
nodeElemCount[i] = 0;
}
Kokkos::View<Index_t*, Kokkos::HostSpace> nodeElemCount("nodeElemCount", numNode());
auto h_nodelist = Kokkos::create_mirror_view(m_nodelist);
Kokkos::deep_copy(h_nodelist, m_nodelist);
for(Index_t i = 0; i < numElem(); ++i)
{
Index_t* nl = nodelist(i);
for(Index_t j = 0; j < 8; ++j)
{
++(nodeElemCount[nl[j]]);
++(nodeElemCount[h_nodelist(i, j)]);
}
}
m_nodeElemStart = Allocate<Index_t>(numNode() + 1);
m_nodeElemStart = Kokkos::View<Index_t*>("m_nodeElemStart", numNode() + 1);
auto h_nodeElemStart = Kokkos::create_mirror_view(m_nodeElemStart);
m_nodeElemStart[0] = 0;
h_nodeElemStart[0] = 0;
for(Index_t i = 1; i <= numNode(); ++i)
{
m_nodeElemStart[i] = m_nodeElemStart[i - 1] + nodeElemCount[i - 1];
h_nodeElemStart[i] = h_nodeElemStart[i - 1] + nodeElemCount[i - 1];
}
m_nodeElemCornerList = Allocate<Index_t>(m_nodeElemStart[numNode()]);
m_nodeElemCornerList =
Kokkos::View<Index_t*>("nodeElemCornerList", h_nodeElemStart[numNode()]);
auto h_nodeElemCornerList = Kokkos::create_mirror_view(m_nodeElemCornerList);
for(Index_t i = 0; i < numNode(); ++i)
{
@@ -407,21 +408,20 @@ Domain::SetupThreadSupportStructures()
for(Index_t i = 0; i < numElem(); ++i)
{
Index_t* nl = nodelist(i);
for(Index_t j = 0; j < 8; ++j)
{
Index_t m = nl[j];
Index_t m = h_nodelist(i, j);
Index_t k = i * 8 + j;
Index_t offset = m_nodeElemStart[m] + nodeElemCount[m];
m_nodeElemCornerList[offset] = k;
Index_t offset = h_nodeElemStart[m] + nodeElemCount[m];
h_nodeElemCornerList[offset] = k;
++(nodeElemCount[m]);
}
}
Index_t clSize = m_nodeElemStart[numNode()];
Index_t clSize = h_nodeElemStart[numNode()];
for(Index_t i = 0; i < clSize; ++i)
{
Index_t clv = m_nodeElemCornerList[i];
Index_t clv = h_nodeElemCornerList[i];
if((clv < 0) || (clv > numElem() * 8))
{
fprintf(
@@ -434,8 +434,8 @@ Domain::SetupThreadSupportStructures()
#endif
}
}
Release<Index_t>(&nodeElemCount);
Kokkos::deep_copy(m_nodeElemCornerList, h_nodeElemCornerList);
Kokkos::deep_copy(m_nodeElemStart, h_nodeElemStart);
}
////////////////////////////////////////////////////////////////////////////////
@@ -486,12 +486,9 @@ Domain::SetupCommBuffers(Int_t edgeNodes)
#endif
// Boundary nodesets
if(m_colLoc == 0)
m_symmX.resize(edgeNodes * edgeNodes);
if(m_rowLoc == 0)
m_symmY.resize(edgeNodes * edgeNodes);
if(m_planeLoc == 0)
m_symmZ.resize(edgeNodes * edgeNodes);
if(m_colLoc == 0) Kokkos::resize(m_symmX, edgeNodes * edgeNodes);
if(m_rowLoc == 0) Kokkos::resize(m_symmY, edgeNodes * edgeNodes);
if(m_planeLoc == 0) Kokkos::resize(m_symmZ, edgeNodes * edgeNodes);
}
////////////////////////////////////////////////////////////////////////////////
@@ -506,9 +503,14 @@ Domain::CreateRegionIndexSets(Int_t nr, Int_t balance)
srand(0);
Index_t myRank = 0;
#endif
this->numReg() = nr;
m_regElemSize = Allocate<Index_t>(numReg());
m_regElemlist = Allocate<Index_t*>(numReg());
this->numReg() = nr;
m_regElemSize = Allocate<Index_t>(numReg());
auto row_map = Kokkos::View<Index_t*>("regElemlist::row_map", numReg() + 1);
auto h_row_map = Kokkos::create_mirror_view(row_map);
auto entries = Kokkos::View<Index_t*>("regElemlist::entries", numElem());
m_regElemlist = t_regElemlist(entries, row_map);
auto h_regElemlist = typename t_regElemlist::HostMirror(
Kokkos::create_mirror_view(m_regElemlist.entries), h_row_map);
Index_t nextIndex = 0;
// if we only have one region just fill it
// Fill out the regNumList with material numbers, which are always
@@ -525,14 +527,14 @@ Domain::CreateRegionIndexSets(Int_t nr, Int_t balance)
// If we have more than one region distribute the elements.
else
{
Int_t regionNum;
Int_t regionVar;
Int_t lastReg = -1;
Int_t binSize;
Index_t elements;
Index_t runto = 0;
Int_t costDenominator = 0;
Int_t* regBinEnd = Allocate<Int_t>(numReg());
Int_t regionNum;
Int_t regionVar;
Int_t lastReg = -1;
Int_t binSize;
Index_t elements;
Index_t runto = 0;
Int_t costDenominator = 0;
Kokkos::View<Int_t*, Kokkos::HostSpace> regBinEnd("regBinEnd", numReg());
// Determine the relative weights of all the regions. This is based off the -b
// flag. Balance is the value passed into b.
for(Index_t i = 0; i < numReg(); ++i)
@@ -612,23 +614,28 @@ Domain::CreateRegionIndexSets(Int_t nr, Int_t balance)
// Second, allocate each region index set
for(Index_t i = 0; i < numReg(); ++i)
{
m_regElemlist[i] = Allocate<Int_t>(regElemSize(i));
h_row_map(i + 1) = regElemSize(i);
regElemSize(i) = 0;
}
// Third, fill index sets
for(Index_t i = 0; i < numElem(); ++i)
{
Index_t r = regNumList(i) - 1; // region index == regnum-1
Index_t regndx = regElemSize(r)++; // Note increment
regElemlist(r, regndx) = i;
Index_t r = regNumList(i) - 1; // region index == regnum-1
Index_t regndx = regElemSize(r)++; // Note increment
h_regElemlist.entries(h_row_map(r) + regndx) = i;
}
Kokkos::deep_copy(m_regElemlist.entries, h_regElemlist.entries);
Kokkos::deep_copy(row_map, h_row_map);
}
/////////////////////////////////////////////////////////////
void
Domain::SetupSymmetryPlanes(Int_t edgeNodes)
{
Index_t nidx = 0;
Index_t nidx = 0;
auto h_symmZ = Kokkos::create_mirror_view(m_symmZ);
auto h_symmY = Kokkos::create_mirror_view(m_symmY);
auto h_symmX = Kokkos::create_mirror_view(m_symmX);
for(Index_t i = 0; i < edgeNodes; ++i)
{
Index_t planeInc = i * edgeNodes * edgeNodes;
@@ -637,55 +644,72 @@ Domain::SetupSymmetryPlanes(Int_t edgeNodes)
{
if(m_planeLoc == 0)
{
m_symmZ[nidx] = rowInc + j;
h_symmZ[nidx] = rowInc + j;
}
if(m_rowLoc == 0)
{
m_symmY[nidx] = planeInc + j;
h_symmY[nidx] = planeInc + j;
}
if(m_colLoc == 0)
{
m_symmX[nidx] = planeInc + j * edgeNodes;
h_symmX[nidx] = planeInc + j * edgeNodes;
}
++nidx;
}
}
Kokkos::deep_copy(m_symmZ, h_symmZ);
Kokkos::deep_copy(m_symmY, h_symmY);
Kokkos::deep_copy(m_symmX, h_symmX);
}
/////////////////////////////////////////////////////////////
void
Domain::SetupElementConnectivities(Int_t edgeElems)
{
lxim(0) = 0;
// With C++17 we wouldn't need to do this and could run this on the GPU
// using class lambdas
auto h_lxim = Kokkos::create_mirror_view(m_lxim);
auto h_lxip = Kokkos::create_mirror_view(m_lxip);
h_lxim(0) = 0;
for(Index_t i = 1; i < numElem(); ++i)
{
lxim(i) = i - 1;
lxip(i - 1) = i;
h_lxim(i) = i - 1;
h_lxip(i - 1) = i;
}
lxip(numElem() - 1) = numElem() - 1;
h_lxip(numElem() - 1) = numElem() - 1;
Kokkos::deep_copy(m_lxim, h_lxim);
Kokkos::deep_copy(m_lxip, h_lxip);
auto h_letam = Kokkos::create_mirror_view(m_letam);
auto h_letap = Kokkos::create_mirror_view(m_letap);
for(Index_t i = 0; i < edgeElems; ++i)
{
letam(i) = i;
letap(numElem() - edgeElems + i) = numElem() - edgeElems + i;
h_letam(i) = i;
h_letap(numElem() - edgeElems + i) = numElem() - edgeElems + i;
}
for(Index_t i = edgeElems; i < numElem(); ++i)
{
letam(i) = i - edgeElems;
letap(i - edgeElems) = i;
h_letam(i) = i - edgeElems;
h_letap(i - edgeElems) = i;
}
Kokkos::deep_copy(m_letam, h_letam);
Kokkos::deep_copy(m_letap, h_letap);
auto h_lzetam = Kokkos::create_mirror_view(m_lzetam);
auto h_lzetap = Kokkos::create_mirror_view(m_lzetap);
for(Index_t i = 0; i < edgeElems * edgeElems; ++i)
{
lzetam(i) = i;
lzetap(numElem() - edgeElems * edgeElems + i) =
h_lzetam(i) = i;
h_lzetap(numElem() - edgeElems * edgeElems + i) =
numElem() - edgeElems * edgeElems + i;
}
for(Index_t i = edgeElems * edgeElems; i < numElem(); ++i)
{
lzetam(i) = i - edgeElems * edgeElems;
lzetap(i - edgeElems * edgeElems) = i;
h_lzetam(i) = i - edgeElems * edgeElems;
h_lzetap(i - edgeElems * edgeElems) = i;
}
Kokkos::deep_copy(m_lzetam, h_lzetam);
Kokkos::deep_copy(m_lzetap, h_lzetap);
}
/////////////////////////////////////////////////////////////
@@ -693,11 +717,24 @@ void
Domain::SetupBoundaryConditions(Int_t edgeElems)
{
Index_t ghostIdx[6]; // offsets to ghost locations
auto h_elemBC = Kokkos::create_mirror_view(m_elemBC);
auto h_lzetam = Kokkos::create_mirror_view(m_lzetam);
auto h_lzetap = Kokkos::create_mirror_view(m_lzetap);
auto h_letam = Kokkos::create_mirror_view(m_letam);
auto h_letap = Kokkos::create_mirror_view(m_letap);
auto h_lxim = Kokkos::create_mirror_view(m_lxim);
auto h_lxip = Kokkos::create_mirror_view(m_lxip);
Kokkos::deep_copy(h_lzetam, m_lzetam);
Kokkos::deep_copy(h_lzetap, m_lzetap);
Kokkos::deep_copy(h_letam, m_letam);
Kokkos::deep_copy(h_letap, m_letap);
Kokkos::deep_copy(h_lxim, m_lxim);
Kokkos::deep_copy(h_lxip, m_lxip);
// set up boundary condition information
for(Index_t i = 0; i < numElem(); ++i)
{
elemBC(i) = Int_t(0);
h_elemBC(i) = Int_t(0);
}
for(Index_t i = 0; i < 6; ++i)
@@ -750,67 +787,75 @@ Domain::SetupBoundaryConditions(Int_t edgeElems)
{
if(m_planeLoc == 0)
{
elemBC(rowInc + j) |= ZETA_M_SYMM;
h_elemBC(rowInc + j) |= ZETA_M_SYMM;
}
else
{
elemBC(rowInc + j) |= ZETA_M_COMM;
lzetam(rowInc + j) = ghostIdx[0] + rowInc + j;
h_elemBC(rowInc + j) |= ZETA_M_COMM;
h_lzetam(rowInc + j) = ghostIdx[0] + rowInc + j;
}
if(m_planeLoc == m_tp - 1)
{
elemBC(rowInc + j + numElem() - edgeElems * edgeElems) |= ZETA_P_FREE;
h_elemBC(rowInc + j + numElem() - edgeElems * edgeElems) |= ZETA_P_FREE;
}
else
{
elemBC(rowInc + j + numElem() - edgeElems * edgeElems) |= ZETA_P_COMM;
lzetap(rowInc + j + numElem() - edgeElems * edgeElems) =
h_elemBC(rowInc + j + numElem() - edgeElems * edgeElems) |= ZETA_P_COMM;
h_lzetap(rowInc + j + numElem() - edgeElems * edgeElems) =
ghostIdx[1] + rowInc + j;
}
if(m_rowLoc == 0)
{
elemBC(planeInc + j) |= ETA_M_SYMM;
h_elemBC(planeInc + j) |= ETA_M_SYMM;
}
else
{
elemBC(planeInc + j) |= ETA_M_COMM;
letam(planeInc + j) = ghostIdx[2] + rowInc + j;
h_elemBC(planeInc + j) |= ETA_M_COMM;
h_letam(planeInc + j) = ghostIdx[2] + rowInc + j;
}
if(m_rowLoc == m_tp - 1)
{
elemBC(planeInc + j + edgeElems * edgeElems - edgeElems) |= ETA_P_FREE;
h_elemBC(planeInc + j + edgeElems * edgeElems - edgeElems) |= ETA_P_FREE;
}
else
{
elemBC(planeInc + j + edgeElems * edgeElems - edgeElems) |= ETA_P_COMM;
letap(planeInc + j + edgeElems * edgeElems - edgeElems) =
h_elemBC(planeInc + j + edgeElems * edgeElems - edgeElems) |= ETA_P_COMM;
h_letap(planeInc + j + edgeElems * edgeElems - edgeElems) =
ghostIdx[3] + rowInc + j;
}
if(m_colLoc == 0)
{
elemBC(planeInc + j * edgeElems) |= XI_M_SYMM;
h_elemBC(planeInc + j * edgeElems) |= XI_M_SYMM;
}
else
{
elemBC(planeInc + j * edgeElems) |= XI_M_COMM;
lxim(planeInc + j * edgeElems) = ghostIdx[4] + rowInc + j;
h_elemBC(planeInc + j * edgeElems) |= XI_M_COMM;
h_lxim(planeInc + j * edgeElems) = ghostIdx[4] + rowInc + j;
}
if(m_colLoc == m_tp - 1)
{
elemBC(planeInc + j * edgeElems + edgeElems - 1) |= XI_P_FREE;
h_elemBC(planeInc + j * edgeElems + edgeElems - 1) |= XI_P_FREE;
}
else
{
elemBC(planeInc + j * edgeElems + edgeElems - 1) |= XI_P_COMM;
lxip(planeInc + j * edgeElems + edgeElems - 1) = ghostIdx[5] + rowInc + j;
h_elemBC(planeInc + j * edgeElems + edgeElems - 1) |= XI_P_COMM;
h_lxip(planeInc + j * edgeElems + edgeElems - 1) =
ghostIdx[5] + rowInc + j;
}
}
}
Kokkos::deep_copy(m_elemBC, h_elemBC);
Kokkos::deep_copy(m_lzetam, h_lzetam);
Kokkos::deep_copy(m_lzetap, h_lzetap);
Kokkos::deep_copy(m_letam, h_letam);
Kokkos::deep_copy(m_letap, h_letap);
Kokkos::deep_copy(m_lxim, h_lxim);
Kokkos::deep_copy(m_lxip, h_lxip);
}
///////////////////////////////////////////////////////////////////////////
projects/rocprofiler-systems/examples/lulesh/lulesh-util.cc
+8 -9
View File
@@ -15,8 +15,7 @@ StrToInt(const char* token, int* retVal)
char* endptr;
const int decimal_base = 10;
if(token == NULL)
return 0;
if(token == NULL) return 0;
c = token;
*retVal = (int) strtol(c, &endptr, decimal_base);
@@ -229,11 +228,13 @@ VerifyAndWriteFinalOutput(Real_t elapsed_time, Domain& locDom, Int_t nx, Int_t n
((elapsed_time * 1e6) / locDom.cycle()) / (nx * nx * nx * numRanks);
Index_t ElemId = 0;
auto h_e = Kokkos::create_mirror_view(locDom.e_view());
Kokkos::deep_copy(h_e, locDom.e_view());
printf("Run completed: \n");
printf(" Problem size = %i \n", nx);
printf(" MPI tasks = %i \n", numRanks);
printf(" Iteration count = %i \n", locDom.cycle());
printf(" Final Origin Energy = %12.6e \n", locDom.e(ElemId));
printf(" Final Origin Energy = %12.6e \n", h_e(ElemId));
Real_t MaxAbsDiff = Real_t(0.0);
Real_t TotalAbsDiff = Real_t(0.0);
@@ -243,16 +244,14 @@ VerifyAndWriteFinalOutput(Real_t elapsed_time, Domain& locDom, Int_t nx, Int_t n
{
for(Index_t k = j + 1; k < nx; ++k)
{
Real_t AbsDiff = FABS(locDom.e(j * nx + k) - locDom.e(k * nx + j));
Real_t AbsDiff = FABS(h_e(j * nx + k) - h_e(k * nx + j));
TotalAbsDiff += AbsDiff;
if(MaxAbsDiff < AbsDiff)
MaxAbsDiff = AbsDiff;
if(MaxAbsDiff < AbsDiff) MaxAbsDiff = AbsDiff;
Real_t RelDiff = AbsDiff / locDom.e(k * nx + j);
Real_t RelDiff = AbsDiff / h_e(k * nx + j);
if(MaxRelDiff < RelDiff)
MaxRelDiff = RelDiff;
if(MaxRelDiff < RelDiff) MaxRelDiff = RelDiff;
}
}
projects/rocprofiler-systems/examples/lulesh/lulesh-viz.cc
+1 -2
View File
@@ -403,8 +403,7 @@ static void
LULESH_PMPIO_Close(void* file, void* udata)
{
DBfile* db = (DBfile*) file;
if(db)
DBClose(db);
if(db) DBClose(db);
}
# endif
projects/rocprofiler-systems/examples/lulesh/lulesh.cc
+413 -445
View File
File diff suppressed because it is too large Load Diff
projects/rocprofiler-systems/examples/lulesh/lulesh.h
+600 -615
View File
File diff suppressed because it is too large Load Diff
projects/rocprofiler-systems/examples/lulesh/lulesh_tuple.h
+412 -453
View File
File diff suppressed because it is too large Load Diff
projects/rocprofiler-systems/examples/transpose/CMakeLists.txt
+1
View File
@@ -3,6 +3,7 @@ cmake_minimum_required(VERSION 3.15 FATAL_ERROR)
project(omnitrace-transpose LANGUAGES CXX)
find_program(HIPCC_EXECUTABLE NAMES hipcc)
mark_as_advanced(HIPCC_EXECUTABLE)
if(NOT HIPCC_EXECUTABLE)
message(AUTHOR_WARNING "hipcc could not be found. Cannot build transpose target")