Add support for architectures not supported by xsimd (#262)

Author: xhochy

.ci_support/linux_aarch64_python3.10_default.____cpython.yaml

@@ -0,0 +1,16 @@
+BUILD:
+- aarch64-conda_cos7-linux-gnu
+c_compiler:
+- gcc
+c_compiler_version:
+- '12'
+cxx_compiler:
+- gxx
+cxx_compiler_version:
+- '12'
+numpy:
+- '1.21'
+python:
+- 3.10.* *_cpython
+target_platform:
+- linux-aarch64

.ci_support/linux_ppc64le_python3.10_default.____cpython.yaml

@@ -0,0 +1,14 @@
+c_compiler:
+- gcc
+c_compiler_version:
+- '12'
+cxx_compiler:
+- gxx
+cxx_compiler_version:
+- '12'
+numpy:
+- '1.21'
+python:
+- 3.10.* *_cpython
+target_platform:
+- linux-ppc64le

.github/workflows/ci.yml

@@ -48,6 +48,8 @@ jobs:
       matrix:
         CONDA_BUILD_YML:
           - linux_64_python3.7_default.____cpython
+          - linux_aarch64_python3.10_default.____cpython
+          - linux_ppc64le_python3.10_default.____cpython
     steps:
       - name: Pull image
         run: docker pull condaforge/mambaforge:latest

.github/workflows/conda-build.sh

@@ -7,5 +7,10 @@ export CONDA_BUILD_YML=$1
 SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 source ${SCRIPT_DIR}/base.sh $*
 conda activate base
-mamba install -y conda-build
-conda build -m .ci_support/${CONDA_BUILD_YML}.yaml conda.recipe
+mamba install -y boa
+
+if grep -q "linux-aarch64\|linux-ppc64le" .ci_support/${CONDA_BUILD_YML}.yaml; then
+    CONDA_BUILD_ARGS="${CONDA_BUILD_ARGS:-} --no-test"
+fi
+
+conda mambabuild ${CONDA_BUILD_ARGS:-} -m .ci_support/${CONDA_BUILD_YML}.yaml conda.recipe

CHANGELOG.rst

@@ -7,6 +7,13 @@
 Changelog
 =========
 
+unreleased
+----------
+
+**Other changes:**
+
+- Support building on architectures that are unsupported by xsimd.
+
 3.1.8 - 2023-06-13
 ------------------
 

src/tabmat/ext/dense_helpers-tmpl.cpp

@@ -45,7 +45,11 @@ namespace xs = xsimd;
         // setup simd accumulators
         % for ir in range(IBLOCK):
             % for jr in range(JBLOCK):
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+		auto accumsimd${ir}_${jr} = (F)0.0;
+#else
                 auto accumsimd${ir}_${jr} = xs::XSIMD_BROADCAST(((F)0.0));
+#endif
             % endfor
         % endfor
 
@@ -78,10 +82,18 @@ namespace xs = xsimd;
             % endfor
             ) {
             % for ir in range(IBLOCK):
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+		auto Xtd${ir} = *Lptr${ir};
+#else
                 auto Xtd${ir} = xs::load_aligned(Lptr${ir});
+#endif
                 % for jr in range(JBLOCK):
                 {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+		    auto Xsimd = *Rptr${jr};
+#else
                     auto Xsimd = xs::load_aligned(Rptr${jr});
+#endif
                     accumsimd${ir}_${jr} = xs::fma(Xtd${ir}, Xsimd, accumsimd${ir}_${jr});
                 }
                 % endfor
@@ -91,7 +103,11 @@ namespace xs = xsimd;
         // horizontal sum of the simd blocks
         % for ir in range(IBLOCK):
             % for jr in range(JBLOCK):
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+		F accum${ir}_${jr} = accumsimd${ir}_${jr};
+#else
                 F accum${ir}_${jr} = xs::XSIMD_REDUCE_ADD(accumsimd${ir}_${jr});
+#endif
             % endfor
         % endfor
 
@@ -150,7 +166,11 @@ void dense_base${kparallel}(F* R, F* L, F* d, F* out,
                 Py_ssize_t jmin2, Py_ssize_t jmax2, 
                 Py_ssize_t kmin, Py_ssize_t kmax, Int innerblock, Int kstep) 
 {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+    constexpr std::size_t simd_size = 1;
+#else
     constexpr std::size_t simd_size = xsimd::simd_type<F>::size;
+#endif
     for (Py_ssize_t imin = imin2; imin < imax2; imin+=innerblock) {
         Py_ssize_t imax = imin + innerblock; 
         if (imax > imax2) {
@@ -248,7 +268,11 @@ template <typename Int, typename F>
 void _dense${order}_sandwich(Int* rows, Int* cols, F* X, F* d, F* out,
         Int in_n, Int out_m, Int m, Int n, Int thresh1d, Int kratio, Int innerblock) 
 {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+    constexpr std::size_t simd_size = 1;
+#else
     constexpr std::size_t simd_size = xsimd::simd_type<F>::size;
+#endif
     constexpr auto alignment = simd_size * sizeof(F);
 
     bool kparallel = (in_n / (kratio*thresh1d)) > (out_m / thresh1d);
@@ -292,7 +316,11 @@ template <typename Int, typename F>
 void _dense${order}_rmatvec(Int* rows, Int* cols, F* X, F* v, F* out,
         Int n_rows, Int n_cols, Int m, Int n) 
 {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+    constexpr std::size_t simd_size = 1;
+#else
     constexpr std::size_t simd_size = xsimd::simd_type<F>::size;
+#endif
     constexpr std::size_t alignment = simd_size * sizeof(F);
 
     auto outglobal = make_aligned_unique<F>(omp_get_max_threads()*n_cols, alignment);

src/tabmat/ext/sparse_helpers-tmpl.cpp

@@ -30,7 +30,12 @@ void _csr_dense${order}_sandwich(
     Int nrows, Int nA_cols, Int nB_cols
     ) 
 {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+    constexpr Int simd_size = 1;
+#else
     constexpr Int simd_size = xsimd::simd_type<F>::size;
+#endif
+
     constexpr auto alignment = simd_size*sizeof(F);
 
     int kblock = 128;
@@ -95,15 +100,28 @@ void _csr_dense${order}_sandwich(
                         }
 
                         F Q = Adata[A_idx];
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+			auto Qsimd = Q;
+#else
                         auto Qsimd = xs::XSIMD_BROADCAST(Q);
+#endif
 
                         Py_ssize_t Cj = Cjj;
                         Py_ssize_t Cjmax2 = Cjj + ((Cjmax - Cjj) / simd_size) * simd_size;
                         for (; Cj < Cjmax2; Cj+=simd_size) {
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+                            auto Bsimd = R[(Py_ssize_t) (Ck-Ckk) * jblock + (Cj-Cjj)];
+                            auto outsimd = outtemp.get()[Ci * nB_cols_rounded + Cj];
+#else
                             auto Bsimd = xs::load_aligned(&R[(Py_ssize_t) (Ck-Ckk) * jblock + (Cj-Cjj)]);
                             auto outsimd = xs::load_aligned(&outtemp.get()[Ci * nB_cols_rounded + Cj]);
+#endif
                             outsimd = xs::fma(Qsimd, Bsimd, outsimd);
+#ifdef XSIMD_NO_SUPPORTED_ARCHITECTURE
+                            outtemp.get()[Ci * nB_cols_rounded + Cj] = outsimd;
+#else
                             outsimd.store_aligned(&outtemp.get()[Ci * nB_cols_rounded + Cj]);
+#endif
                         }
 
                         for (; Cj < Cjmax; Cj++) {