backend_gpu/field_storage_backend.h

#ifndef GPU_BACKEND_H
#define GPU_BACKEND_H
 
#include "../defs.h"
#include "../field_storage.h"
 
/* CUDA / HIP implementations */
template <typename T>
void field_storage<T>::allocate_field(const lattice_struct &lattice) {
    // Allocate space for the field of the device
    gpuMalloc(&fieldbuf, sizeof(T) * lattice.field_alloc_size());
    if (fieldbuf == nullptr) {
        std::cout << "Failure in field memory allocation\n";
    }
    assert(fieldbuf != nullptr);
}
 
template <typename T>
void field_storage<T>::free_field() {
    if (fieldbuf != nullptr) {
        gpuFree(fieldbuf);
    }
    fieldbuf = nullptr;
}
 
// Only attempt to compile with CUDA compiler.
// Hilapp will skip these.
// #if defined(__CUDACC__) || defined(__HIPCC__)
#if !defined(HILAPP)
 
// These are used in device code. Can be called directly in a kernel.
template <typename T>
__device__ inline auto field_storage<T>::get(const unsigned i,
                                             const unsigned field_alloc_size) const {
    assert(i < field_alloc_size);
    using base_t = hila::arithmetic_type<T>;
    constexpr unsigned n_elements = sizeof(T) / sizeof(base_t);
    T value;
    base_t *value_f = (base_t *)&value;
    base_t *fp = (base_t *)(fieldbuf);
    for (unsigned e = 0; e < n_elements; e++) {
        value_f[e] = fp[e * field_alloc_size + i];
    }
    return value;
}
 
template <typename T>
// template <typename A>
__device__ inline void field_storage<T>::set(const T &value, const unsigned i,
                                             const unsigned field_alloc_size) {
    assert(i < field_alloc_size);
    using base_t = hila::arithmetic_type<T>;
    constexpr unsigned n_elements = sizeof(T) / sizeof(base_t);
    const base_t *value_f = (base_t *)&value;
    base_t *fp = (base_t *)(fieldbuf);
    for (unsigned e = 0; e < n_elements; e++) {
        fp[e * field_alloc_size + i] = value_f[e];
    }
}
 
/// Get a single element from the field outside a loop. Slow, should only be used for
/// setup
template <typename T>
__global__ void get_element_kernel(field_storage<T> field, char *buffer, unsigned i,
                                   const unsigned field_alloc_size) {
    *((T *)buffer) = field.get(i, field_alloc_size);
}
 
template <typename T>
auto field_storage<T>::get_element(const unsigned i, const lattice_struct &lattice) const {
    char *d_buffer;
    T value;
 
    // Call the kernel to collect the element
    gpuMalloc(&(d_buffer), sizeof(T));
    get_element_kernel<<<1, 1>>>(*this, d_buffer, i, lattice.field_alloc_size());
 
    // Copy the result to the host
    gpuMemcpy((char *)(&value), d_buffer, sizeof(T), gpuMemcpyDeviceToHost);
    gpuFree(d_buffer);
    return value;
}
 
/// Set a single element from outside a loop. Slow, should only be used for setup
template <typename T>
__global__ void set_element_kernel(field_storage<T> field, T value, unsigned i,
                                   const unsigned field_alloc_size) {
    field.set(value, i, field_alloc_size);
}
 
template <typename T>
__global__ void set_element_kernel_ptr(field_storage<T> field, const T *buf, unsigned i,
                                       const unsigned field_alloc_size) {
    field.set(*buf, i, field_alloc_size);
}
 
 
template <typename T>
template <typename A>
void field_storage<T>::set_element(A &value, const unsigned i, const lattice_struct &lattice) {
    T t_value = value;
 
    if constexpr (sizeof(T) <= GPU_GLOBAL_ARG_MAX_SIZE) {
        // pass element directly as arg
        set_element_kernel<<<1, 1>>>(*this, t_value, i, lattice.field_alloc_size());
    } else {
        // This branch is used for large variables:
        // allocate space and copy the buffer to the device
        T *d_buffer;
        gpuMalloc(&(d_buffer), sizeof(T));
        gpuMemcpy(d_buffer, (char *)&t_value, sizeof(T), gpuMemcpyHostToDevice);
 
        // call the kernel to set correct indexes
        set_element_kernel_ptr<<<1, 1>>>(*this, d_buffer, i, lattice.field_alloc_size());
        gpuFree(d_buffer);
    }
}
 
/// A kernel that gathers elements
template <typename T>
__global__ void gather_elements_kernel(field_storage<T> field, T *buffer, unsigned *site_index,
                                       const int n, const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        buffer[Index] = field.get(site_index[Index], field_alloc_size);
    }
}
 
/// CUDA implementation of gather_elements without CUDA aware MPI
template <typename T>
void field_storage<T>::gather_elements(T *RESTRICT buffer, const unsigned *RESTRICT index_list,
                                       int n, const lattice_struct &lattice) const {
    unsigned *d_site_index;
    T *d_buffer;
 
    // Copy the list of boundary site indexes to the device
    gpuMalloc(&(d_site_index), n * sizeof(unsigned));
    gpuMemcpy(d_site_index, index_list, n * sizeof(unsigned), gpuMemcpyHostToDevice);
 
    // Call the kernel to build the list of elements
    gpuMalloc(&(d_buffer), n * sizeof(T));
    int N_blocks = n / N_threads + 1;
    gather_elements_kernel<<<N_blocks, N_threads>>>(*this, d_buffer, d_site_index, n,
                                                    lattice.field_alloc_size());
 
    // Copy the result to the host
    gpuMemcpy((char *)buffer, d_buffer, n * sizeof(T), gpuMemcpyDeviceToHost);
 
    gpuFree(d_site_index);
    gpuFree(d_buffer);
}
 
/// A kernel that gathers elements negated
// requires unary -
template <typename T, std::enable_if_t<hila::has_unary_minus<T>::value, int> = 0>
__global__ void gather_elements_negated_kernel(field_storage<T> field, T *buffer,
                                               unsigned *site_index, const int n,
                                               const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        buffer[Index] = -field.get(site_index[Index], field_alloc_size);
    }
}
 
/// CUDA implementation of gather_elements_negated without CUDA aware MPI
// The only difference to the above is the kernel that is called
template <typename T>
void field_storage<T>::gather_elements_negated(T *RESTRICT buffer,
                                               const unsigned *RESTRICT index_list, int n,
                                               const lattice_struct &lattice) const {
    unsigned *d_site_index;
    T *d_buffer;
 
    if constexpr (!hila::has_unary_minus<T>::value) {
        assert(sizeof(T) < 0 && "Unary 'operatur- ()' must be defined for Field variable "
                                "for antiperiodic b.c.");
    } else {
 
        // Copy the list of boundary site indexes to the device
        gpuMalloc(&(d_site_index), n * sizeof(unsigned));
        gpuMemcpy(d_site_index, index_list, n * sizeof(unsigned), gpuMemcpyHostToDevice);
 
        // Call the kernel to build the list of elements
        gpuMalloc(&(d_buffer), n * sizeof(T));
        int N_blocks = n / N_threads + 1;
        gather_elements_negated_kernel<<<N_blocks, N_threads>>>(*this, d_buffer, d_site_index, n,
                                                                lattice.field_alloc_size());
 
        // Copy the result to the host
        gpuMemcpy(buffer, d_buffer, n * sizeof(T), gpuMemcpyDeviceToHost);
 
        gpuFree(d_site_index);
        gpuFree(d_buffer);
    }
}
 
template <typename T>
__global__ void gather_comm_elements_kernel(field_storage<T> field, T *buffer, unsigned *site_index,
                                            const int n, const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        using base_t = hila::arithmetic_type<T>;
        constexpr unsigned n_elements = sizeof(T) / sizeof(base_t);
        T element = field.get(site_index[Index], field_alloc_size);
        base_t *ep = (base_t *)&element;
        base_t *fp = (base_t *)(buffer);
        for (unsigned e = 0; e < n_elements; e++) {
            fp[Index + n * e] = ep[e];
        }
    }
}
 
// gather -elements only if unary - exists
template <typename T, std::enable_if_t<hila::has_unary_minus<T>::value, int> = 0>
__global__ void gather_comm_elements_negated_kernel(field_storage<T> field, T *buffer,
                                                    unsigned *site_index, const int n,
                                                    const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        using base_t = hila::arithmetic_type<T>;
        constexpr unsigned n_elements = sizeof(T) / sizeof(base_t);
        T element = -field.get(site_index[Index], field_alloc_size);
        base_t *ep = (base_t *)&element;
        base_t *fp = (base_t *)(buffer);
        for (unsigned e = 0; e < n_elements; e++) {
            fp[Index + n * e] = ep[e];
        }
    }
}
 
// Index list is constant? Map each cpu pointer to a device pointer and copy just once
struct cuda_comm_node_struct {
    const unsigned *cpu_index;
    unsigned *gpu_index;
    int n;
};
 
inline unsigned *get_site_index(const lattice_struct::comm_node_struct &to_node, Parity par,
                                int &n) {
    static std::vector<struct cuda_comm_node_struct> comm_nodes;
 
    const unsigned *cpu_index = to_node.get_sitelist(par, n);
    for (struct cuda_comm_node_struct comm_node : comm_nodes) {
        if (cpu_index == comm_node.cpu_index && n == comm_node.n) {
            return comm_node.gpu_index;
        }
    }
    struct cuda_comm_node_struct comm_node;
    comm_node.cpu_index = cpu_index;
    comm_node.n = n;
    gpuMalloc(&(comm_node.gpu_index), n * sizeof(unsigned));
    gpuMemcpy(comm_node.gpu_index, cpu_index, n * sizeof(unsigned), gpuMemcpyHostToDevice);
    comm_nodes.push_back(comm_node);
    return comm_node.gpu_index;
}
 
// MPI buffer on the device. Use the gather_elements and gather_elements_negated
// kernels to fill the buffer.
template <typename T>
void field_storage<T>::gather_comm_elements(T *buffer,
                                            const lattice_struct::comm_node_struct &to_node,
                                            Parity par, const lattice_struct &lattice,
                                            bool antiperiodic) const {
    int n;
    unsigned *d_site_index = get_site_index(to_node, par, n);
    T *d_buffer;
 
#ifdef GPU_AWARE_MPI
    // The buffer is already on the device
    d_buffer = buffer;
#else
    // Allocate a buffer on the device
    gpuMalloc(&(d_buffer), n * sizeof(T));
#endif
 
    // Call the kernel to build the list of elements
    int N_blocks = n / N_threads + 1;
    if (antiperiodic) {
 
        if constexpr (hila::has_unary_minus<T>::value) {
            gather_comm_elements_negated_kernel<<<N_blocks, N_threads>>>(
                *this, d_buffer, d_site_index, n, lattice.field_alloc_size());
        }
 
    } else {
        gather_comm_elements_kernel<<<N_blocks, N_threads>>>(*this, d_buffer, d_site_index, n,
                                                             lattice.field_alloc_size());
    }
 
#ifndef GPU_AWARE_MPI
    // Copy the result to the host
    gpuMemcpy((char *)buffer, d_buffer, n * sizeof(T), gpuMemcpyDeviceToHost);
    gpuFree(d_buffer);
#endif
}
 
/// A kernel that scatters the elements
template <typename T>
__global__ void place_elements_kernel(field_storage<T> field, T *buffer, unsigned *site_index,
                                      const int n, const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        field.set(buffer[Index], site_index[Index], field_alloc_size);
    }
}
 
/// CUDA implementation of place_elements without CUDA aware MPI
template <typename T>
void field_storage<T>::place_elements(T *RESTRICT buffer, const unsigned *RESTRICT index_list,
                                      int n, const lattice_struct &lattice) {
    unsigned *d_site_index;
    T *d_buffer;
 
    // Allocate space and copy the buffer to the device
    gpuMalloc(&(d_buffer), n * sizeof(T));
    gpuMemcpy(d_buffer, buffer, n * sizeof(T), gpuMemcpyHostToDevice);
 
    // Copy the list of boundary site indexes to the device
    gpuMalloc(&(d_site_index), n * sizeof(unsigned));
    gpuMemcpy(d_site_index, index_list, n * sizeof(unsigned), gpuMemcpyHostToDevice);
 
    // Call the kernel to place the elements
    int N_blocks = n / N_threads + 1;
    place_elements_kernel<<<N_blocks, N_threads>>>(*this, d_buffer, d_site_index, n,
                                                   lattice.field_alloc_size());
 
    gpuFree(d_buffer);
    gpuFree(d_site_index);
}
 
template <typename T, std::enable_if_t<hila::has_unary_minus<T>::value, int> = 0>
__global__ void set_local_boundary_elements_kernel(field_storage<T> field, unsigned offset,
                                                   unsigned *site_index, const int n,
                                                   const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        T value;
        value = -field.get(site_index[Index], field_alloc_size);
        field.set(value, offset + Index, field_alloc_size);
    }
}
 
template <typename T>
void field_storage<T>::set_local_boundary_elements(Direction dir, Parity par,
                                                   const lattice_struct &lattice,
                                                   bool antiperiodic) {
    // Only need to do something for antiperiodic boundaries
#ifdef SPECIAL_BOUNDARY_CONDITIONS
    if (antiperiodic) {
        if constexpr (hila::has_unary_minus<T>::value) {
            unsigned n, start = 0;
            if (par == ODD) {
                n = lattice.special_boundaries[dir].n_odd;
                start = lattice.special_boundaries[dir].n_even;
            } else {
                if (par == EVEN)
                    n = lattice.special_boundaries[dir].n_even;
                else
                    n = lattice.special_boundaries[dir].n_total;
            }
            unsigned offset = lattice.special_boundaries[dir].offset + start;
 
            unsigned *d_site_index;
            check_device_error("earlier");
            gpuMalloc(&d_site_index, n * sizeof(unsigned));
            gpuMemcpy(d_site_index, lattice.special_boundaries[dir].move_index + start,
                      n * sizeof(unsigned), gpuMemcpyHostToDevice);
 
            unsigned N_blocks = n / N_threads + 1;
            set_local_boundary_elements_kernel<<<N_blocks, N_threads>>>(
                *this, offset, d_site_index, n, lattice.field_alloc_size());
 
            gpuFree(d_site_index);
        } else {
            assert("Antiperiodic b.c. cannot be used with unsigned field elements");
        }
    }
 
#else
    assert(!antiperiodic && "antiperiodic only with SPECIAL_BOUNDARY_CONDITIONS defined");
#endif
}
 
// Place communicated elements to the field array
template <typename T>
__global__ void place_comm_elements_kernel(field_storage<T> field, T *buffer, unsigned offset,
                                           const int n, const unsigned field_alloc_size) {
    unsigned Index = threadIdx.x + blockIdx.x * blockDim.x;
    if (Index < n) {
        using base_t = hila::arithmetic_type<T>;
        constexpr unsigned n_elements = sizeof(T) / sizeof(base_t);
        T element;
        base_t *ep = (base_t *)&element;
        base_t *fp = (base_t *)(buffer);
        for (unsigned e = 0; e < n_elements; e++) {
            ep[e] = fp[Index + n * e];
        }
        field.set(element, offset + Index, field_alloc_size);
    }
}
 
// Standard MPI, buffer is on the cpu and needs to be copied accross
template <typename T>
void field_storage<T>::place_comm_elements(Direction d, Parity par, T *buffer,
                                           const lattice_struct::comm_node_struct &from_node,
                                           const lattice_struct &lattice) {
 
    unsigned n = from_node.n_sites(par);
    T *d_buffer;
 
#ifdef GPU_AWARE_MPI
    // MPI buffer is on device
    d_buffer = buffer;
#else
    // Allocate space and copy the buffer to the device
    gpuMalloc(&(d_buffer), n * sizeof(T));
    gpuMemcpy(d_buffer, buffer, n * sizeof(T), gpuMemcpyHostToDevice);
#endif
 
    unsigned N_blocks = n / N_threads + 1;
    place_comm_elements_kernel<<<N_blocks, N_threads>>>((*this), d_buffer, from_node.offset(par), n,
                                                        lattice.field_alloc_size());
 
#ifndef GPU_AWARE_MPI
    gpuFree(d_buffer);
#endif
}
 
#ifdef GPU_AWARE_MPI
 
template <typename T>
void field_storage<T>::free_mpi_buffer(T *d_buffer) {
    gpuFree(d_buffer);
}
 
template <typename T>
T *field_storage<T>::allocate_mpi_buffer(unsigned n) {
    T *d_buffer;
    gpuMalloc(&(d_buffer), n * sizeof(T));
    return d_buffer;
}
 
#else
 
template <typename T>
void field_storage<T>::free_mpi_buffer(T *buffer) {
    std::free(buffer);
}
 
template <typename T>
T *field_storage<T>::allocate_mpi_buffer(unsigned n) {
    return (T *)memalloc(n * sizeof(T));
}
 
#endif
 
#else // now HILAPP
 
// Hilapp requires a dummy implementation for functions with auto return type
template <typename T>
auto field_storage<T>::get(const unsigned i, const unsigned field_alloc_size) const {
    T value;
    return value;
}
template <typename T>
auto field_storage<T>::get_element(const unsigned i, const lattice_struct &lattice) const {
    T value;
    return value;
}
 
#endif // !HILAPP .. HILAPP
 
#endif