Single Precision Support

.github/workflows/ubuntu.yml

@@ -199,6 +199,56 @@ jobs:
       run: |
         python3 -m pytest tests/
 
+  clang_sp:
+    name: Clang@18.0 w/ SP
+    runs-on: ubuntu-24.04
+    env:
+      CXXFLAGS: "-Werror -Wshadow -Woverloaded-virtual -Wextra-semi -Wunreachable-code -fno-operator-names -Wno-pass-failed"
+      LDFLAGS: "-fuse-ld=lld"
+      CXX: "clang++"
+      CC: "clang"
+      AMReX_PRECISION: "SINGLE"
+      AMReX_PARTICLES_PRECISION: "SINGLE"
+    steps:
+      - uses: actions/checkout@v4
+      - name: Dependencies
+        run: |
+          .github/workflows/dependencies/dependencies_clang18.sh
+          .github/workflows/dependencies/dependencies_ccache.sh
+      - name: Set Up Cache
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/ccache
+          key: ccache-${{ github.workflow }}-${{ github.job }}-git-${{ github.sha }}
+          restore-keys: |
+            ccache-${{ github.workflow }}-${{ github.job }}-git-
+      - name: Build & Install
+        run: |
+          export CCACHE_COMPRESS=1
+          export CCACHE_COMPRESSLEVEL=10
+          export CCACHE_MAXSIZE=300M
+          ccache -z
+
+          export CMAKE_BUILD_PARALLEL_LEVEL=4
+
+          export CC=$(which clang-18)
+          export CXX=$(which clang++-18)
+          python3 -m pip install -U pip
+          python3 -m pip install -U build packaging setuptools[core] wheel
+          python3 -m pip install -U cmake
+          python3 -m pip install -U pandas pytest
+          python3 -m pip install -v .
+          python3 -c "import amrex.space1d as amr; print(amr.__version__)"
+          python3 -c "import amrex.space2d as amr; print(amr.__version__)"
+          python3 -c "import amrex.space3d as amr; print(amr.__version__)"
+
+          ccache -s
+          du -hs ~/.cache/ccache
+
+      - name: Unit tests
+        run: |
+          python3 -m pytest tests/
+
   nvcc11:
     name: CUDA@12.2 GNU@10.5
     runs-on: ubuntu-24.04

setup.py

@@ -95,7 +95,7 @@ def build_extension(self, ext):
             "-DAMReX_GPU_BACKEND=" + AMReX_GPU_BACKEND,
             "-DAMReX_MPI:BOOL=" + AMReX_MPI,
             "-DAMReX_PRECISION=" + AMReX_PRECISION,
-            #'-DAMReX_PARTICLES_PRECISION=' + AMReX_PARTICLES_PRECISION,
+            "-DAMReX_PARTICLES_PRECISION=" + AMReX_PARTICLES_PRECISION,
             "-DpyAMReX_CCACHE=" + PYAMREX_CCACHE,
             "-DpyAMReX_IPO=" + PYAMREX_IPO,
             ## dependency control (developers & package managers)
@@ -174,6 +174,7 @@ def build_extension(self, ext):
 AMReX_GPU_BACKEND = os.environ.get("AMREX_GPU_BACKEND", "NONE")
 AMReX_MPI = os.environ.get("AMREX_MPI", "OFF")
 AMReX_PRECISION = os.environ.get("AMREX_PRECISION", "DOUBLE")
+AMReX_PARTICLES_PRECISION = os.environ.get("AMREX_PARTICLES_PRECISION", "DOUBLE")
 #   single value or as a list 1;2;3
 AMReX_SPACEDIM = os.environ.get("AMREX_SPACEDIM", "1;2;3")
 BUILD_SHARED_LIBS = os.environ.get("AMREX_BUILD_SHARED_LIBS", "OFF")

src/Base/AMReX.cpp

@@ -82,6 +82,24 @@ void init_AMReX(py::module& m)
                 return "SYCL";
 #else
                 return py::none();
+#endif
+            })
+        .def_property_readonly_static(
+            "precision",
+            [](py::object){
+#ifdef AMREX_SINGLE_PRECISION
+                return "SINGLE";
+#else
+                return "DOUBLE";
+#endif
+        })
+        .def_property_readonly_static(
+            "precision_particles",
+            [](py::object){
+#ifdef AMREX_SINGLE_PRECISION_PARTICLES
+                return "SINGLE";
+#else
+                return "DOUBLE";
 #endif
             })
         ;

src/Particle/ArrayOfStructs.H

@@ -43,17 +43,38 @@ namespace
 #if (AMREX_SPACEDIM >= 3)
         descr.append(py::make_tuple("z", py::format_descriptor<RealType>::format()));
 #endif
+
+        /* Note AMReX_Particle.H:
+         * struct alignas(sizeof(double)) Particle ...
+         *   aligns the ParticleBase member (pads the end if needed).
+         * in
+         * struct ParticleBase
+         * {
+         *   T m_pos[AMREX_SPACEDIM];
+         *   T m_rdata[NReal];
+         *   uint64_t m_idcpu = 0;
+         *   int m_idata[NInt];
+         * };
+         */
         if constexpr (ParticleType::NReal > 0) {
             for(int ii=0; ii < ParticleType::NReal; ii++) {
                 descr.append(py::make_tuple("rdata_"+std::to_string(ii),py::format_descriptor<RealType>::format()));
             }
         }
+#ifdef AMREX_SINGLE_PRECISION_PARTICLES
+        if constexpr ((AMREX_SPACEDIM + ParticleType::NReal) % 2 != 0) {  // alignas
+            descr.append(py::make_tuple("", "|V4"));  // empty space due to alignment
+        }
+#endif
         descr.append(py::make_tuple("idcpu", py::format_descriptor<uint64_t>::format()) );
         if constexpr (ParticleType::NInt > 0) {
             for(int ii=0; ii < ParticleType::NInt; ++ii) {
                 descr.append(py::make_tuple("idata_"+std::to_string(ii),py::format_descriptor<int>::format()));
             }
-        }
+        }/*
+        if constexpr (ParticleType::NInt % 2 != 0) {  // alignas
+            descr.append(py::make_tuple("", "|V4"));  // empty space due to alignment
+        }*/
 
         d["descr"] = descr;
         d["version"] = 3;

tests/test_aos.py

@@ -76,15 +76,16 @@ def test_array_interface():
     # print('particle 2 from aos:\n',aos[1])
     # print('array interface\n', aos.__array_interface__)
     arr = aos.to_numpy()
+    int_arg = 7 if arr[0][0].dtype == "float32" else 6  # padding
     assert (
         np.isclose(arr[0][0], 1.0)
         and np.isclose(arr[0][4], 5.2)
-        and np.isclose(arr[0][6], 6)
+        and np.isclose(arr[0][int_arg], 6)  # fixme in SP: reads int32(0)
     )
     assert (
         np.isclose(arr[1][2], 10)
         and np.isclose(arr[1][3], 11.1)
-        and np.isclose(arr[1][6], 13)
+        and np.isclose(arr[1][int_arg], 13)
     )
 
     p3 = amr.Particle_2_1(x=-3)
@@ -102,10 +103,11 @@ def test_array_interface():
     assert aos[0].y == arr[0][1] == 0
     assert aos[0].z == arr[0][2] == 0
 
-    shape = amr.Config.spacedim + amr.Particle_2_1.NReal + amr.Particle_2_1.NInt + 1
-    for ii in range(shape):
-        arr[1][ii] = 0
-    arr[1][1] = -5  # np.array([0, -5, 0,0,0,0,0])
+    arr[1][0] = 0
+    arr[1][1] = -5
+    arr[1][2] = 0
+    arr[1][3] = 0
+    arr[1][4] = 0  # np.array([0, -5, 0,0,0,0,0, ...])
     print("array:", arr)
     print("aos[0]:", aos[0], "aos[1]:", aos[1])
     assert aos[1].y == arr[1][1] == -5

tests/test_parmparse.py

@@ -1,6 +1,8 @@
 # -*- coding: utf-8 -*-
 import os
 
+import numpy as np
+
 import amrex.space3d as amr
 
 
@@ -14,7 +16,7 @@ def test_parmparse():
     dopml = pp_param.get_bool("do_pml")
 
     assert dopml
-    assert dt == 1.0e-5
+    assert np.isclose(dt, 1.0e-5)
     assert ncell == 100
 
     pp.pretty_print_table()

tests/test_particle.py

@@ -37,19 +37,31 @@ def test_particle_init():
 def test_particle_set():
     p1 = amr.Particle_8_0()
     p1.setPos(1, 1.5)
-    assert p1.pos(0) == 0 and p1.pos(1) == 1.5 and p1.pos(2) == 0
+    assert (
+        np.isclose(p1.pos(0), 0)
+        and np.isclose(p1.pos(1), 1.5)
+        and np.isclose(p1.pos(2), 0)
+    )
     p1.setPos([1.0, 1, 2])
-    assert p1.pos(0) == 1 and p1.pos(1) == 1 and p1.pos(2) == 2
+    assert (
+        np.isclose(p1.pos(0), 1)
+        and np.isclose(p1.pos(1), 1)
+        and np.isclose(p1.pos(2), 2)
+    )
     p1.setPos(amr.RealVect(2, 3.3, 4.2))
-    assert p1.pos(0) == 2 and p1.pos(1) == 3.3 and p1.pos(2) == 4.2
+    assert (
+        np.isclose(p1.pos(0), 2)
+        and np.isclose(p1.pos(1), 3.3)
+        and np.isclose(p1.pos(2), 4.2)
+    )
 
     print(p1.x, p1.y, p1.z)
     p1.x = 2.1
-    assert p1.x == 2.1
+    assert np.isclose(p1.x, 2.1)
     p1.y = 3.2
-    assert p1.y == 3.2
+    assert np.isclose(p1.y, 3.2)
     p1.z = 5.1
-    assert p1.z == 5.1
+    assert np.isclose(p1.z, 5.1)
 
 
 def test_rdata():

tests/test_particleContainer.py

@@ -251,10 +251,10 @@ def test_pc_init():
 
             aos = pti.aos()
             aos_arr = aos.to_numpy()
-            assert aos[0].x == 0.30
-            assert aos[0].y == 0.35
-            assert aos[0].z == 0.40
-            assert aos_arr[0]["z"] == 0.40
+            assert np.isclose(aos[0].x, 0.30)
+            assert np.isclose(aos[0].y, 0.35)
+            assert np.isclose(aos[0].z, 0.40)
+            assert np.isclose(aos_arr[0]["z"], 0.40)
 
             soa = pti.soa()
             real_arrays = soa.get_real_data()
@@ -294,7 +294,9 @@ def test_particle_init(Npart, particle_container):
             assert int_arrays[0][0] == 1
             assert isinstance(aos_arr[0]["rdata_0"], np.floating)
             assert isinstance(aos_arr[0]["idata_0"], np.integer)
-            assert np.isclose(aos_arr[0]["rdata_0"], 0.5) and aos_arr[0]["idata_0"] == 5
+            assert (
+                np.isclose(aos_arr[0]["rdata_0"], 0.5) and aos_arr[0]["idata_0"] == 5
+            )  # fixme in SP: random value np.int32(-2147483648) == 5
 
             aos_arr[0]["idata_0"] = 2
             aos1 = pt.get_array_of_structs()
@@ -337,7 +339,7 @@ def test_particle_init(Npart, particle_container):
         aos = pt.get_array_of_structs()
         print(aos[0])
         assert aos[0].get_idata(0) == 2
-        assert real_arrays[1][0] == -1.2
+        assert np.isclose(real_arrays[1][0], -1.2)
         assert int_arrays[0][0] == -3
 
 
@@ -362,7 +364,7 @@ def test_per_cell(empty_particle_container, std_geometry, std_particle):
         pc.total_number_of_particles() == pc.number_of_particles_at_level(0) == ncells
     )
     print("npts * real_1", ncells * std_particle.real_array_data[1])
-    assert ncells * std_particle.real_array_data[1] == sum_1
+    assert np.isclose(ncells * std_particle.real_array_data[1], sum_1)
 
 
 def test_soa_pc_numpy(soa_particle_container, Npart):
@@ -462,6 +464,10 @@ class Config:
 @pytest.mark.skipif(
     importlib.util.find_spec("pandas") is None, reason="pandas is not available"
 )
+@pytest.mark.skipif(
+    amr.Config.precision_particles == "SINGLE",
+    reason="Requires DOUBLE precision particles",
+)
 def test_pc_df(particle_container, Npart):
     pc = particle_container
     print(f"pc={pc}")
@@ -550,6 +556,10 @@ def test_pc_empty_df(empty_particle_container, Npart):
 @pytest.mark.skipif(
     importlib.util.find_spec("pandas") is None, reason="pandas is not available"
 )
+@pytest.mark.skipif(
+    amr.Config.precision_particles == "SINGLE",
+    reason="Requires DOUBLE precision particles",
+)
 @pytest.mark.skipif(not amr.Config.have_mpi, reason="Requires AMReX_MPI=ON")
 def test_pc_df_mpi(particle_container, Npart):
     pc = particle_container

tests/test_particleTile.py

@@ -111,7 +111,7 @@ def test_ptile_soa():
     ar1 = np.array(rdata[1], copy=False)
     ir0 = np.array(idata[0], copy=False)
     print(ar0.dtype)
-    assert ar0.dtype == "float"
+    assert ar0.dtype == "float" or ar0.dtype == "float32"
     assert ir0.dtype == "int32"
     print("---------")
     ir0[0] = -55

tests/test_realbox.py

@@ -1,5 +1,7 @@
 # -*- coding: utf-8 -*-
 
+import numpy as np
+
 import amrex.space3d as amr
 from amrex.space3d import RealVect as RV
 from amrex.space3d import XDim3
@@ -14,9 +16,9 @@ def test_realbox_empty():
 
     rb.setLo([-1, -2, -3])
     rb.setHi([0, 0, 0])
-    assert rb.volume() == 6
+    assert np.isclose(rb.volume(), 6)
     rb.setHi(0, 1.2)
-    assert rb.hi(0) == 1.2
+    assert np.isclose(rb.hi(0), 1.2)
 
 
 def test_realbox_frombox(std_box):

tests/test_realvect.py

@@ -13,8 +13,8 @@ def test_realvect_init():
 
     for ii in range(amr.Config.spacedim):
         assert rv[ii] == 0
-        assert rv2[ii] == ii + 0.1
-        assert rv3[ii] == 0.3
+        assert np.isclose(rv2[ii], ii + 0.1)
+        assert np.isclose(rv3[ii], 0.3)
 
 
 @pytest.mark.skipif(amr.Config.spacedim != 1, reason="Requires AMREX_SPACEDIM = 1")
@@ -49,8 +49,8 @@ def test_rv_3d1():
     # Check indexing
     assert obj[0] == 1
     assert obj[1] == 2
-    assert obj[2] == 3.14
-    assert obj[-1] == 3.14
+    assert np.isclose(obj[2], 3.14)
+    assert np.isclose(obj[-1], 3.14)
     assert obj[-2] == 2
     assert obj[-3] == 1
     with pytest.raises(IndexError):
@@ -137,9 +137,13 @@ def test_subtraction():
         assert v4[ii] == v5[ii]
 
     # v - v
-    assert v5 - v3 == uv
+    res1 = v5 - v3
+    for ii in range(amr.Config.spacedim):
+        assert np.isclose(res1[ii], uv[ii])
     # r - v
-    assert 1.0 - v6 == amr.RealVect(2.0)
+    res2 = 1.0 - v6
+    for ii in range(amr.Config.spacedim):
+        assert np.isclose(res2[ii], 2.0)
 
 
 def test_multiplication():