ENH: Covariance CPU SPMD support

cpp/oneapi/dal/algo/covariance/backend/cpu/compute_kernel_dense.cpp

@@ -18,10 +18,14 @@
 
 #include "oneapi/dal/algo/covariance/backend/cpu/compute_kernel.hpp"
 #include "oneapi/dal/algo/covariance/backend/cpu/compute_kernel_common.hpp"
+#include "oneapi/dal/algo/covariance/backend/cpu/partial_compute_kernel.hpp"
+#include "oneapi/dal/algo/covariance/backend/cpu/finalize_compute_kernel.hpp"
 #include "oneapi/dal/backend/interop/common.hpp"
 #include "oneapi/dal/backend/interop/error_converter.hpp"
 #include "oneapi/dal/backend/interop/table_conversion.hpp"
 
+#include "oneapi/dal/backend/primitives/utils.hpp"
+
 #include "oneapi/dal/table/row_accessor.hpp"
 
 namespace oneapi::dal::covariance::backend {
@@ -30,13 +34,83 @@ using dal::backend::context_cpu;
 using descriptor_t = detail::descriptor_base<task::compute>;
 using parameters_t = detail::compute_parameters<task::compute>;
 
+namespace be = dal::backend;
+namespace pr = be::primitives;
 namespace daal_covariance = daal::algorithms::covariance;
 namespace interop = dal::backend::interop;
 
 template <typename Float, daal::CpuType Cpu>
 using daal_covariance_kernel_t = daal_covariance::internal::
     CovarianceDenseBatchKernel<Float, daal_covariance::Method::defaultDense, Cpu>;
 
+template <typename Float, typename Task>
+static compute_result<Task> call_daal_spmd_kernel(const context_cpu& ctx,
+                                                  const detail::descriptor_base<Task>& desc,
+                                                  const detail::compute_parameters<Task>& params,
+                                                  const table& data) {
+    auto& comm = ctx.get_communicator();
+    const std::int64_t component_count = data.get_column_count();
+
+    // Compute partial results locally on this rank's data
+    partial_compute_input<Task> partial_input(data);
+    auto partial_result =
+        partial_compute_kernel_cpu<Float, method::by_default, Task>{}(ctx, desc, partial_input);
+
+    // Extract partial results as mutable arrays
+    auto nobs_nd = pr::table2ndarray<Float>(partial_result.get_partial_n_rows());
+    auto sums_nd = pr::table2ndarray<Float>(partial_result.get_partial_sum());
+    auto crossproduct_nd = pr::table2ndarray<Float>(partial_result.get_partial_crossproduct());
+
+    auto nobs_ary = dal::array<Float>::wrap(nobs_nd.get_mutable_data(), nobs_nd.get_count());
+    auto sums_ary = dal::array<Float>::wrap(sums_nd.get_mutable_data(), sums_nd.get_count());
+    auto crossproduct_ary =
+        dal::array<Float>::wrap(crossproduct_nd.get_mutable_data(), crossproduct_nd.get_count());
+
+    // The DAAL online kernel stores centered crossproducts:
+    //   cp = X^T*X - sums*sums^T/nobs
+    // Simple allreduce of centered crossproducts is incorrect because each
+    // rank uses its local mean. Un-center before allreduce, then re-center
+    // with global statistics after.
+    if (!desc.get_assume_centered()) {
+        Float* cp_ptr = crossproduct_ary.get_mutable_data();
+        const Float* sums_ptr = sums_ary.get_data();
+        const Float local_nobs = *nobs_ary.get_data();
+        const Float inv_nobs = Float(1) / local_nobs;
+        for (std::int64_t i = 0; i < component_count; ++i) {
+            for (std::int64_t j = 0; j < component_count; ++j) {
+                cp_ptr[i * component_count + j] += inv_nobs * sums_ptr[i] * sums_ptr[j];
+            }
+        }
+    }
+
+    // Allreduce raw crossproduct, sums, and nobs across all ranks
+    comm.allreduce(nobs_ary).wait();
+    comm.allreduce(sums_ary).wait();
+    comm.allreduce(crossproduct_ary).wait();
+
+    // Re-center with global statistics
+    if (!desc.get_assume_centered()) {
+        Float* cp_ptr = crossproduct_ary.get_mutable_data();
+        const Float* sums_ptr = sums_ary.get_data();
+        const Float global_nobs = *nobs_ary.get_data();
+        const Float inv_nobs = Float(1) / global_nobs;
+        for (std::int64_t i = 0; i < component_count; ++i) {
+            for (std::int64_t j = 0; j < component_count; ++j) {
+                cp_ptr[i * component_count + j] -= inv_nobs * sums_ptr[i] * sums_ptr[j];
+            }
+        }
+    }
+
+    // Reconstruct aggregated partial result and finalize
+    partial_compute_result<Task> aggregated;
+    aggregated.set_partial_n_rows(homogen_table::wrap(nobs_ary, 1, 1));
+    aggregated.set_partial_sum(homogen_table::wrap(sums_ary, 1, component_count));
+    aggregated.set_partial_crossproduct(
+        homogen_table::wrap(crossproduct_ary, component_count, component_count));
+
+    return finalize_compute_kernel_cpu<Float, method::by_default, Task>{}(ctx, desc, aggregated);
+}
+
 template <typename Float, typename Task>
 static compute_result<Task> call_daal_kernel(const context_cpu& ctx,
                                              const detail::descriptor_base<Task>& desc,
@@ -121,6 +195,9 @@ static compute_result<Task> compute(const context_cpu& ctx,
                                     const detail::descriptor_base<Task>& desc,
                                     const detail::compute_parameters<Task>& params,
                                     const compute_input<Task>& input) {
+    if (ctx.get_communicator().get_rank_count() > 1) {
+        return call_daal_spmd_kernel<Float, Task>(ctx, desc, params, input.get_data());
+    }
     return call_daal_kernel<Float, Task>(ctx, desc, params, input.get_data());
 }
 

cpp/oneapi/dal/algo/covariance/compute_types.hpp

@@ -118,6 +118,8 @@ class compute_input : public base {
     /// property value
     compute_input(const table& data);
 
+    virtual ~compute_input() = default;
+
     /// An $n \\times p$ table with the training data, where each row stores one
     /// feature vector.
     /// @remark default = table{}
@@ -249,6 +251,8 @@ class partial_compute_input : protected compute_input<Task> {
 
     partial_compute_input(const partial_compute_result<Task>& prev, const table& data);
 
+    virtual ~partial_compute_input() = default;
+
     const table& get_data() const {
         return compute_input<Task>::get_data();
     }

cpp/oneapi/dal/algo/covariance/detail/compute_ops.cpp

@@ -34,7 +34,7 @@ struct compute_ops_dispatcher<Policy, Float, Method, Task> {
     compute_parameters<Task> select_parameters(const Policy& ctx,
                                                const descriptor_base<Task>& desc,
                                                const compute_input<Task>& input) const {
-        using kernel_dispatcher_t = dal::backend::kernel_dispatcher<KERNEL_SINGLE_NODE_CPU(
+        using kernel_dispatcher_t = dal::backend::kernel_dispatcher<KERNEL_UNIVERSAL_SPMD_CPU(
             parameters::compute_parameters_cpu<Float, Method, Task>)>;
         return kernel_dispatcher_t{}(ctx, desc, input);
     }
@@ -51,7 +51,7 @@ struct compute_ops_dispatcher<Policy, Float, Method, Task> {
                                const descriptor_base<Task>& desc,
                                const compute_parameters<Task>& params,
                                const compute_input<Task>& input) const {
-        using kernel_dispatcher_t = dal::backend::kernel_dispatcher<KERNEL_SINGLE_NODE_CPU(
+        using kernel_dispatcher_t = dal::backend::kernel_dispatcher<KERNEL_UNIVERSAL_SPMD_CPU(
             backend::compute_kernel_cpu<Float, Method, Task>)>;
         return kernel_dispatcher_t()(ctx, desc, params, input);
     }

cpp/oneapi/dal/algo/covariance/test/spmd.cpp

@@ -84,7 +84,6 @@ TEMPLATE_LIST_TEST_M(covariance_spmd_test,
                      "covariance common flow",
                      "[covariance][integration][spmd]",
                      covariance_types) {
-    SKIP_IF(this->get_policy().is_cpu());
     SKIP_IF(this->not_float64_friendly());
 
     using Float = std::tuple_element_t<0, TestType>;