Merge branch 'develop' into collected-small-changes

lib/kokkos/Makefile.kokkos

@@ -23,7 +23,7 @@ KOKKOS_DEVICES ?= "OpenMP"
 # NVIDIA:   Kepler,Kepler30,Kepler32,Kepler35,Kepler37,Maxwell,Maxwell50,Maxwell52,Maxwell53,Pascal60,Pascal61,Volta70,Volta72,Turing75,Ampere80,Ampere86,Ada89,Hopper90
 # ARM:      ARMv80,ARMv81,ARMv8-ThunderX,ARMv8-TX2,A64FX,ARMv9-Grace
 # IBM:      Power8,Power9
-# AMD-GPUS: AMD_GFX906,AMD_GFX908,AMD_GFX90A,AMD_GFX940,AMD_GFX942,AMD_GFX1030,AMD_GFX1100
+# AMD-GPUS: AMD_GFX906,AMD_GFX908,AMD_GFX90A,AMD_GFX940,AMD_GFX942,AMD_GFX1030,AMD_GFX1100,AMD_GFX1103
 # AMD-CPUS: AMDAVX,Zen,Zen2,Zen3
 # Intel-GPUs: Intel_Gen,Intel_Gen9,Intel_Gen11,Intel_Gen12LP,Intel_DG1,Intel_XeHP,Intel_PVC
 KOKKOS_ARCH ?= ""
@@ -465,6 +465,7 @@ KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100 := $(call kokkos_has_string,$(KOKKOS_ARCH),
 ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100), 0)
   KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100 := $(call kokkos_has_string,$(KOKKOS_ARCH),NAVI1100)
 endif
+KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1103 := $(call kokkos_has_string,$(KOKKOS_ARCH),AMD_GFX1103)
 
 # Any AVX?
 KOKKOS_INTERNAL_USE_ARCH_AVX        := $(shell expr $(KOKKOS_INTERNAL_USE_ARCH_SNB) + $(KOKKOS_INTERNAL_USE_ARCH_AMDAVX))
@@ -1158,6 +1159,11 @@ ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100), 1)
   tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GPU")
   KOKKOS_INTERNAL_HIP_ARCH_FLAG := --offload-arch=gfx1100
 endif
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1103), 1)
+  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GFX1103")
+  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GPU")
+  KOKKOS_INTERNAL_HIP_ARCH_FLAG := --offload-arch=gfx1103
+endif
 
 
 ifeq ($(KOKKOS_INTERNAL_USE_HIP), 1)

lib/kokkos/cmake/KokkosCore_config.h.in

@@ -120,6 +120,7 @@
 #cmakedefine KOKKOS_ARCH_AMD_GFX942
 #cmakedefine KOKKOS_ARCH_AMD_GFX1030
 #cmakedefine KOKKOS_ARCH_AMD_GFX1100
+#cmakedefine KOKKOS_ARCH_AMD_GFX1103
 #cmakedefine KOKKOS_ARCH_AMD_GPU
 #cmakedefine KOKKOS_ARCH_VEGA // deprecated
 #cmakedefine KOKKOS_ARCH_VEGA906 // deprecated

lib/kokkos/cmake/kokkos_arch.cmake

@@ -103,9 +103,9 @@ LIST(APPEND CORRESPONDING_AMD_FLAGS  gfx90a   gfx90a      gfx908   gfx908)
 LIST(APPEND SUPPORTED_AMD_GPUS       MI50/60  MI50/60)
 LIST(APPEND SUPPORTED_AMD_ARCHS      VEGA906  AMD_GFX906)
 LIST(APPEND CORRESPONDING_AMD_FLAGS  gfx906   gfx906)
-LIST(APPEND SUPPORTED_AMD_GPUS       RX7900XTX  RX7900XTX    V620/W6800  V620/W6800)
-LIST(APPEND SUPPORTED_AMD_ARCHS      NAVI1100   AMD_GFX1100  NAVI1030    AMD_GFX1030)
-LIST(APPEND CORRESPONDING_AMD_FLAGS  gfx1100    gfx1100      gfx1030     gfx1030)
+LIST(APPEND SUPPORTED_AMD_GPUS       PHOENIX      RX7900XTX    V620/W6800  V620/W6800)
+LIST(APPEND SUPPORTED_AMD_ARCHS      AMD_GFX1103  AMD_GFX1100  NAVI1030    AMD_GFX1030)
+LIST(APPEND CORRESPONDING_AMD_FLAGS  gfx1103      gfx1100      gfx1030     gfx1030)
 
 #FIXME CAN BE REPLACED WITH LIST_ZIP IN CMAKE 3.17
 FOREACH(ARCH IN LISTS SUPPORTED_AMD_ARCHS)

lib/kokkos/core/src/HIP/Kokkos_HIP_Instance.hpp

@@ -35,7 +35,8 @@ struct HIPTraits {
   static constexpr int WarpSize       = 64;
   static constexpr int WarpIndexMask  = 0x003f; /* hexadecimal for 63 */
   static constexpr int WarpIndexShift = 6;      /* WarpSize == 1 << WarpShift*/
-#elif defined(KOKKOS_ARCH_AMD_GFX1030) || defined(KOKKOS_ARCH_AMD_GFX1100)
+#elif defined(KOKKOS_ARCH_AMD_GFX1030) || defined(KOKKOS_ARCH_AMD_GFX1100) || \
+    defined(KOKKOS_ARCH_AMD_GFX1103)
   static constexpr int WarpSize       = 32;
   static constexpr int WarpIndexMask  = 0x001f; /* hexadecimal for 31 */
   static constexpr int WarpIndexShift = 5;      /* WarpSize == 1 << WarpShift*/

lib/kokkos/core/src/OpenMPTarget/Kokkos_OpenMPTarget_ParallelScan_Range.hpp

@@ -143,7 +143,7 @@ class ParallelScan<FunctorType, Kokkos::RangePolicy<Traits...>,
             local_offset_value = element_values(team_id, i - 1);
             // FIXME_OPENMPTARGET We seem to access memory illegaly on AMD GPUs
 #if defined(KOKKOS_ARCH_AMD_GPU) && !defined(KOKKOS_ARCH_AMD_GFX1030) && \
-    !defined(KOKKOS_ARCH_AMD_GFX1100)
+    !defined(KOKKOS_ARCH_AMD_GFX1100) && !defined(KOKKOS_ARCH_AMD_GFX1103)
             if constexpr (Analysis::Reducer::has_join_member_function()) {
               if constexpr (std::is_void_v<WorkTag>)
                 a_functor_reducer.get_functor().join(local_offset_value,

lib/kokkos/core/src/impl/Kokkos_Core.cpp

@@ -752,6 +752,9 @@ void pre_initialize_internal(const Kokkos::InitializationSettings& settings) {
 #elif defined(KOKKOS_ARCH_AMD_GFX1100)
   declare_configuration_metadata("architecture", "GPU architecture",
                                  "AMD_GFX1100");
+#elif defined(KOKKOS_ARCH_AMD_GFX1103)
+  declare_configuration_metadata("architecture", "GPU architecture",
+                                 "AMD_GFX1103");
 
 #else
   declare_configuration_metadata("architecture", "GPU architecture", "none");

lib/kokkos/generate_makefile.bash

@@ -164,6 +164,7 @@ display_help_text() {
       echo "                 AMD_GFX942      = AMD GPU MI300 GFX942"
       echo "                 AMD_GFX1030     = AMD GPU V620/W6800 GFX1030"
       echo "                 AMD_GFX1100     = AMD GPU RX 7900 XT(X) GFX1100"
+      echo "                 AMD_GFX1103     = AMD APU Radeon 740M/760M/780M/880M/890M GFX1103"
       echo "               [ARM]"
       echo "                 ARMV80          = ARMv8.0 Compatible CPU"
       echo "                 ARMV81          = ARMv8.1 Compatible CPU"

src/KOKKOS/domain_kokkos.cpp

@@ -563,7 +563,6 @@ void DomainKokkos::operator()(TagDomain_image_flip, const int &i) const {
 void DomainKokkos::lamda2x(int n)
 {
   atomKK->sync(Device,X_MASK);
-
   x = atomKK->k_x.view<LMPDeviceType>();
 
   copymode = 1;
@@ -573,6 +572,21 @@ void DomainKokkos::lamda2x(int n)
   atomKK->modified(Device,X_MASK);
 }
 
+void DomainKokkos::lamda2x(int n, int groupbit_in)
+{
+  atomKK->sync(Device,X_MASK);
+  x = atomKK->k_x.view<LMPDeviceType>();
+  mask = atomKK->k_mask.view<LMPDeviceType>();
+  mask = atomKK->k_mask.view<LMPDeviceType>();
+  groupbit = groupbit_in;
+
+  copymode = 1;
+  Kokkos::parallel_for(Kokkos::RangePolicy<LMPDeviceType, TagDomain_lamda2x_group>(0,n),*this);
+  copymode = 0;
+
+  atomKK->modified(Device,X_MASK);
+}
+
 KOKKOS_INLINE_FUNCTION
 void DomainKokkos::operator()(TagDomain_lamda2x, const int &i) const {
   const double xi1 = x(i,1);
@@ -582,6 +596,17 @@ void DomainKokkos::operator()(TagDomain_lamda2x, const int &i) const {
   x(i,2) = h[2]*xi2 + boxlo[2];
 }
 
+KOKKOS_INLINE_FUNCTION
+void DomainKokkos::operator()(TagDomain_lamda2x_group, const int &i) const {
+  if (mask[i] & groupbit) {
+    const double xi1 = x(i,1);
+    const double xi2 = x(i,2);
+    x(i,0) = h[0]*x(i,0) + h[5]*xi1 + h[4]*xi2 + boxlo[0];
+    x(i,1) = h[1]*xi1 + h[3]*xi2 + boxlo[1];
+    x(i,2) = h[2]*xi2 + boxlo[2];
+  }
+}
+
 /* ----------------------------------------------------------------------
    convert box coords to triclinic 0-1 lamda coords for all N atoms
    lamda = H^-1 (x - x0)
@@ -590,7 +615,6 @@ void DomainKokkos::operator()(TagDomain_lamda2x, const int &i) const {
 void DomainKokkos::x2lamda(int n)
 {
   atomKK->sync(Device,X_MASK);
-
   x = atomKK->k_x.view<LMPDeviceType>();
 
   copymode = 1;
@@ -600,6 +624,20 @@ void DomainKokkos::x2lamda(int n)
   atomKK->modified(Device,X_MASK);
 }
 
+void DomainKokkos::x2lamda(int n, int groupbit_in)
+{
+  atomKK->sync(Device,X_MASK);
+  x = atomKK->k_x.view<LMPDeviceType>();
+  mask = atomKK->k_mask.view<LMPDeviceType>();
+  groupbit = groupbit_in;
+
+  copymode = 1;
+  Kokkos::parallel_for(Kokkos::RangePolicy<LMPDeviceType, TagDomain_x2lamda_group>(0,n),*this);
+  copymode = 0;
+
+  atomKK->modified(Device,X_MASK);
+}
+
 KOKKOS_INLINE_FUNCTION
 void DomainKokkos::operator()(TagDomain_x2lamda, const int &i) const {
   F_FLOAT delta[3];
@@ -612,3 +650,17 @@ void DomainKokkos::operator()(TagDomain_x2lamda, const int &i) const {
   x(i,2) = h_inv[2]*delta[2];
 }
 
+KOKKOS_INLINE_FUNCTION
+void DomainKokkos::operator()(TagDomain_x2lamda_group, const int &i) const {
+  if (mask[i] & groupbit) {
+    F_FLOAT delta[3];
+    delta[0] = x(i,0) - boxlo[0];
+    delta[1] = x(i,1) - boxlo[1];
+    delta[2] = x(i,2) - boxlo[2];
+
+    x(i,0) = h_inv[0]*delta[0] + h_inv[5]*delta[1] + h_inv[4]*delta[2];
+    x(i,1) = h_inv[1]*delta[1] + h_inv[3]*delta[2];
+    x(i,2) = h_inv[2]*delta[2];
+  }
+}
+

src/KOKKOS/domain_kokkos.h

@@ -24,7 +24,9 @@ namespace LAMMPS_NS {
 struct TagDomain_remap_all{};
 struct TagDomain_image_flip{};
 struct TagDomain_lamda2x{};
+struct TagDomain_lamda2x_group{};
 struct TagDomain_x2lamda{};
+struct TagDomain_x2lamda_group{};
 
 class DomainKokkos : public Domain {
  public:
@@ -35,7 +37,9 @@ class DomainKokkos : public Domain {
   void remap_all();
   void image_flip(int, int, int);
   void x2lamda(int) override;
+  void x2lamda(int,int) override;
   void lamda2x(int) override;
+  void lamda2x(int,int) override;
   // forward remaining x2lamda() and lambda2x() variants to parent class
   void x2lamda(double *a, double *b) override { Domain::x2lamda(a,b); }
   void lamda2x(double *a, double *b) override { Domain::lamda2x(a,b); }
@@ -54,18 +58,26 @@ class DomainKokkos : public Domain {
   KOKKOS_INLINE_FUNCTION
   void operator()(TagDomain_lamda2x, const int&) const;
 
+  KOKKOS_INLINE_FUNCTION
+  void operator()(TagDomain_lamda2x_group, const int&) const;
+
   KOKKOS_INLINE_FUNCTION
   void operator()(TagDomain_x2lamda, const int&) const;
 
+  KOKKOS_INLINE_FUNCTION
+  void operator()(TagDomain_x2lamda_group, const int&) const;
+
   static KOKKOS_INLINE_FUNCTION
   Few<double,3> unmap(Few<double,3> prd, Few<double,6> h, int triclinic,
       Few<double,3> x, imageint image);
 
  private:
+  int groupbit;
   double lo[3],hi[3],period[3];
   int n_flip, m_flip, p_flip;
   ArrayTypes<LMPDeviceType>::t_x_array x;
   ArrayTypes<LMPDeviceType>::t_imageint_1d image;
+  ArrayTypes<LMPDeviceType>::t_int_1d mask;
 };
 
 KOKKOS_INLINE_FUNCTION