outer.hpp

Go to the documentation of this file.

//=======================================================================
// Copyright (c) 2014-2023 Baptiste Wicht
// Distributed under the terms of the MIT License.
// (See accompanying file LICENSE or copy at
//  http://opensource.org/licenses/MIT)
//=======================================================================

#pragma once

#ifdef ETL_BLAS_MODE
#include "cblas.h"
#endif

namespace etl::impl::blas {

#ifdef ETL_BLAS_MODE

template <typename A, typename B, typename C>
void outer(const A& a, const B& b, C&& c) {
    c = 0;

    a.ensure_cpu_up_to_date();
    b.ensure_cpu_up_to_date();
    c.ensure_cpu_up_to_date();

    if constexpr (all_single_precision<A, B, C>) {
        cblas_sger(CblasRowMajor, etl::dim<0>(a), etl::dim<0>(b), 1.0, a.memory_start(), 1, b.memory_start(), 1, c.memory_start(), etl::dim<0>(b));
    } else {
        cblas_dger(CblasRowMajor, etl::dim<0>(a), etl::dim<0>(b), 1.0, a.memory_start(), 1, b.memory_start(), 1, c.memory_start(), etl::dim<0>(b));
    }

    c.invalidate_gpu();
}

template <typename A, typename B, typename C>
void batch_outer(const A& a, const B& b, C&& c) {
    const size_t m = etl::rows(c);
    const size_t n = etl::columns(c);
    const size_t k = etl::rows(a);

    a.ensure_cpu_up_to_date();
    b.ensure_cpu_up_to_date();

    if constexpr (all_single_precision<A, B, C>) {
        cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, m, n, k, 1.0f, a.memory_start(), m, b.memory_start(), n, 0.0f, c.memory_start(), n);
    } else {
        cblas_dgemm(CblasRowMajor, CblasTrans, CblasNoTrans, m, n, k, 1.0, a.memory_start(), m, b.memory_start(), n, 0.0, c.memory_start(), n);
    }

    c.invalidate_gpu();
}

#else

template <typename A, typename B, typename C>
void outer(const A& /*a*/, const B& /*b*/, C&& /*c*/) {
    cpp_unreachable("BLAS not enabled/available");
}

template <typename A, typename B, typename C>
void batch_outer(const A& /*a*/, const B& /*b*/, C&& /*c*/) {
    cpp_unreachable("BLAS not enabled/available");
}

#endif

} //end of namespace etl::impl::blas