kernel_launch.hpp

Go to the documentation of this file.

#pragma once
#ifndef MULTI_WRAPPER_IMPLS_LAUNCH_HPP_
#define MULTI_WRAPPER_IMPLS_LAUNCH_HPP_

#include "kernel.hpp"
#include "../types.hpp"
#include "../memory.hpp"
#include "../stream.hpp"
#include "../kernel_launch.hpp"
#include "../pointer.hpp"
#include "../device.hpp"

// The following is needed for occupancy-related calculation convenience functions
#include <cuda_runtime.h>

namespace cuda {

template<typename Kernel, typename... KernelParameters>
void enqueue_launch(
    Kernel&&                kernel,
    const stream_t&         stream,
    launch_configuration_t  launch_configuration,
    KernelParameters&&...   parameters)
{
    static_assert(
        detail_::all_true<::std::is_trivially_copy_constructible<detail_::kernel_parameter_decay_t<KernelParameters>>::value...>::value,
        "All kernel parameter types must be of a trivially copy-constructible (decayed) type." );
    static constexpr const bool wrapped_contextual_kernel = ::std::is_base_of<kernel_t, typename ::std::decay<Kernel>::type>::value;
#if CUDA_VERSION >= 12000
    static constexpr const bool library_kernel = cuda::detail_::is_library_kernel<Kernel>::value;
#else
    static constexpr const bool library_kernel = false;
#endif // CUDA_VERSION >= 12000
#ifndef NDEBUG
    // wrapped kernel and library kernel compatibility with the launch configuration
    // will be validated further inside, when we differentiate them from raw kernels
    detail_::validate(launch_configuration);
#endif

    // We would have liked an "if constexpr" here, but that is unsupported by C++11, so we have to
    // use tagged dispatch for the separate behavior for raw and wrapped kernels - although the enqueue_launch
    // function for each of them will basically be just a one-liner :-(
    detail_::enqueue_launch<Kernel, KernelParameters...>(
        detail_::bool_constant<wrapped_contextual_kernel>{},
        detail_::bool_constant<library_kernel>{},
        ::std::forward<Kernel>(kernel), stream, launch_configuration,
        ::std::forward<KernelParameters>(parameters)...);
}

namespace detail_ {

inline void validate_shared_mem_compatibility(
    const device_t &device,
    memory::shared::size_t shared_mem_size) noexcept(false)
{
    if (shared_mem_size == 0) { return; }
    memory::shared::size_t max_shared = device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN);

    // Note: A single kernel may not be able to access this shared memory capacity without opting-in to
    // it using kernel_t::set_maximum_dynamic_shared_memory_per_block. See @ref kernel_t

    if (shared_mem_size > max_shared) {
        throw ::std::invalid_argument(
            "A dynamic shared memory size of " + ::std::to_string(shared_mem_size)
            + " bytes exceeds the device maximum of " + ::std::to_string(max_shared));
    }
}

inline void validate_compatibility(
    const device::id_t            device_id,
    memory::shared::size_t        shared_mem_size,
    bool                          cooperative_launch,
    optional<grid::dimensions_t>  block_cluster_dimensions) noexcept(false)
{
    auto device = device::get(device_id);
    if (not cooperative_launch or device.supports_block_cooperation()) {
        throw ::std::runtime_error(device::detail_::identify(device_id)
            + " cannot launch kernels with inter-block cooperation");
    }
    validate_shared_mem_compatibility(device, shared_mem_size);
    if (block_cluster_dimensions) {
#if CUDA_VERSION >= 12000
        if (not device.supports_block_clustering()) {
            throw ::std::runtime_error(device::detail_::identify(device_id)
                + " cannot launch kernels with inter-block cooperation");
            // TODO: Uncomment this once the CUDA driver offers info on the maximum
            // cluster size...
            //
            // auto max_cluster_size = ???;
            // auto cluster_size = block_cluster_dimensions.value().volume();
            // if (cluster_size > max_cluster_size) {
            //  throw ::std::runtime_error(device::detail_::identify(device_id)
            //      + " only supports as many as " + ::std::to_string(max_cluster_size)
            //      + "blocks per block-cluster, but " + ::std::to_string(cluster_size));
        }
#else
        throw ::std::runtime_error("Block clusters are not supported with CUDA versions earlier than 12.0");
#endif // CUDA_VERSION >= 12000
    }

    // The CUDA driver does not offer us information with which we could check the validity
    // of trying a programmatically dependent launch, or a programmatic completion event,
    // or the use of a "remote" memory synchronization domain. So, assuming that's all valid
}

template <typename Dims>
inline void validate_any_dimensions_compatibility(
    const device_t &device, Dims dims,
    Dims maxima,
    const char* kind) noexcept(false)
{
    auto device_id = device.id();
    auto check =
        [device_id, kind](grid::dimension_t dim, grid::dimension_t max, const char *axis) {
            if (max < dim) {
                throw ::std::invalid_argument(
                    ::std::string("specified ") + kind + " " + axis + "-axis dimension " + ::std::to_string(dim)
                    + " exceeds the maximum supported " + axis + " dimension of " + ::std::to_string(max)
                    + " for " + device::detail_::identify(device_id));
            }
        };
    check(dims.x, maxima.x, "X");
    check(dims.y, maxima.y, "Y");
    check(dims.z, maxima.z, "Z");
}

inline void validate_block_dimension_compatibility(
    const device_t &device,
    grid::block_dimensions_t block_dims) noexcept(false)
{
    auto max_block_size = device.maximum_threads_per_block();
    auto volume = block_dims.volume();
    if (volume > max_block_size) {
        throw ::std::invalid_argument(
            "Specified block dimensions result in blocks of size " + ::std::to_string(volume)
            + ", exceeding the maximum possible block size of " + ::std::to_string(max_block_size)
            + " for " + device::detail_::identify(device.id()));
    }
    auto maxima = grid::block_dimensions_t{
        static_cast<grid::block_dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X)),
        static_cast<grid::block_dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y)),
        static_cast<grid::block_dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z))
    };
    validate_any_dimensions_compatibility(device, block_dims, maxima, "block");
}

inline void validate_grid_dimension_compatibility(
    const device_t &device,
    grid::block_dimensions_t block_dims) noexcept(false)
{
    auto maxima = grid::dimensions_t{
        static_cast<grid::dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X)),
        static_cast<grid::dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y)),
        static_cast<grid::dimension_t>(device.get_attribute(CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z))
    };
    validate_any_dimensions_compatibility(device, block_dims, maxima, "grid");
}


inline void validate_shared_mem_size_compatibility(
    const kernel_t& kernel_ptr,
    memory::shared::size_t shared_mem_size)
{
    if (shared_mem_size == 0) { return; }
    auto max_shared = kernel_ptr.get_maximum_dynamic_shared_memory_per_block();
    if (shared_mem_size > max_shared) {
        throw ::std::invalid_argument(
            "Requested dynamic shared memory size "
            + ::std::to_string(shared_mem_size) + " exceeds kernel's maximum allowed value of "
            + ::std::to_string(max_shared));
    }
}

inline void validate_block_dimension_compatibility(
    const kernel_t&          kernel,
    grid::block_dimensions_t block_dims)
{
    auto max_block_size = kernel.maximum_threads_per_block();
    auto volume = block_dims.volume();
    if (volume > max_block_size) {
        throw ::std::invalid_argument(
            "specified block dimensions result in blocks of size " + ::std::to_string(volume)
            + ", exceeding the maximum possible block size of " + ::std::to_string(max_block_size)
            + " for " + kernel::detail_::identify(kernel));
    }
}

inline void validate_dyanmic_shared_memory_size(
    const kernel_t&          kernel,
    memory::shared::size_t   dynamic_shared_memory_size)
{
    memory::shared::size_t max_dyn_shmem = kernel.get_attribute(
        kernel::attribute_t::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES);
    if (dynamic_shared_memory_size > max_dyn_shmem) {
        throw ::std::invalid_argument(
            "specified size of dynamic shared memory, " + ::std::to_string(dynamic_shared_memory_size)
            + "bytes, exceeds the maximum supported by  " + kernel::detail_::identify(kernel)
            + ", " + ::std::to_string(max_dyn_shmem) + " bytes");
    }
}


template<typename... KernelParameters>
void enqueue_launch_helper<kernel::apriori_compiled_t, KernelParameters...>::operator()(
    const kernel::apriori_compiled_t&  wrapped_kernel,
    const stream_t &                  stream,
    launch_configuration_t            launch_configuration,
    KernelParameters &&...            parameters) const
{
    using raw_kernel_t = typename kernel::detail_::raw_kernel_typegen<KernelParameters ...>::type;
    auto unwrapped_kernel_function = reinterpret_cast<raw_kernel_t>(const_cast<void *>(wrapped_kernel.ptr()));
    // Notes:
    // 1. The inner cast here is because we store the pointer as const void* - as an extra
    //    precaution against anybody trying to write through it. Now, function pointers
    //    can't get written through, but are still for some reason not considered const.
    // 2. We rely on the caller providing us with more-or-less the correct parameters -
    //    corresponding to the compiled kernel function's. I say "more or less" because the
    //    `KernelParameter` pack may contain some references, arrays and so on - which CUDA
    //    kernels cannot accept; so we massage those a bit.

    // It is assumed arguments were already been validated

    detail_::enqueue_raw_kernel_launch_in_current_context(
        unwrapped_kernel_function,
        stream.device_id(),
        stream.context_handle(),
        stream.handle(),
        launch_configuration,
        ::std::forward<KernelParameters>(parameters)...);
}

template<typename... KernelParameters>
::std::array<const void*, sizeof...(KernelParameters)>
marshal_dynamic_kernel_arguments(KernelParameters&&... parameters)
{
    return ::std::array<const void*, sizeof...(KernelParameters)> { &parameters... };
}

// Note: The last (valid) element of marshalled_arguments must be null
inline void enqueue_kernel_launch_by_handle_in_current_context(
    kernel::handle_t        kernel_function_handle,
    device::id_t            device_id,
    context::handle_t       context_handle,
    stream::handle_t        stream_handle,
    launch_configuration_t  launch_config,
    const void**            marshalled_arguments)
{
    // It is assumed arguments were already been validated

    status_t status;
    const auto&lc = launch_config; // alias for brevity
#if CUDA_VERSION >= 12000
    CUlaunchAttribute launch_attributes[detail_::maximum_possible_kernel_launch_attributes+1];
    auto launch_attributes_span = span<CUlaunchAttribute>{
        launch_attributes, sizeof(launch_attributes)/sizeof(launch_attributes[0])
    };
    CUlaunchConfig full_launch_config = detail_::marshal(lc, stream_handle, launch_attributes_span);
    status = cuLaunchKernelEx(
        &full_launch_config,
        kernel_function_handle,
        const_cast<void**>(marshalled_arguments),
        nullptr);
#else
    if (launch_config.has_nondefault_attributes())
        status = cuLaunchCooperativeKernel(
            kernel_function_handle,
            lc.dimensions.grid.x,  lc.dimensions.grid.y,  lc.dimensions.grid.z,
            lc.dimensions.block.x, lc.dimensions.block.y, lc.dimensions.block.z,
            lc.dynamic_shared_memory_size,
            stream_handle,
            const_cast<void**>(marshalled_arguments)
        );
    else {
        static constexpr const auto no_arguments_in_alternative_format = nullptr;
        // TODO: Consider passing marshalled_arguments in the alternative format
        status = cuLaunchKernel(
            kernel_function_handle,
            lc.dimensions.grid.x,  lc.dimensions.grid.y,  lc.dimensions.grid.z,
            lc.dimensions.block.x, lc.dimensions.block.y, lc.dimensions.block.z,
            lc.dynamic_shared_memory_size,
            stream_handle,
            const_cast<void**>(marshalled_arguments),
            no_arguments_in_alternative_format
        );
    }
#endif // CUDA_VERSION >= 12000
    throw_if_error_lazy(status,
        ::std::string(" kernel launch failed for ") + kernel::detail_::identify(kernel_function_handle)
        + " on " + stream::detail_::identify(stream_handle, context_handle, device_id));
}


template<typename... KernelParameters>
struct enqueue_launch_helper<kernel_t, KernelParameters...> {

    void operator()(
    const kernel_t&                       wrapped_kernel,
    const stream_t&                       stream,
    launch_configuration_t                launch_config,
    KernelParameters&&...                 arguments) const
    {
    // It is assumed arguments were already been validated

#ifndef NDEBUG
        if (wrapped_kernel.context() != stream.context()) {
            throw ::std::invalid_argument{"Attempt to launch " + kernel::detail_::identify(wrapped_kernel)
                + " on " + stream::detail_::identify(stream) + ": Different contexts"};
        }
        validate_compatibility(wrapped_kernel, launch_config);
#endif
        auto marshalled_arguments { marshal_dynamic_kernel_arguments(::std::forward<KernelParameters>(arguments)...) };
        auto function_handle = wrapped_kernel.handle();
        CAW_SET_SCOPE_CONTEXT(stream.context_handle());

        enqueue_kernel_launch_by_handle_in_current_context(
            function_handle, stream.device_id(), stream.context_handle(),
            stream.handle(), launch_config, marshalled_arguments.data());
    } // operator()
};

template<typename RawKernelFunction, typename... KernelParameters>
void enqueue_launch(
    bool_constant<false>, // Not a wrapped contextual kernel,
    bool_constant<false>, // and not a library kernel, so it must be a raw kernel function
    RawKernelFunction&&       kernel_function,
    const stream_t&           stream,
    launch_configuration_t    launch_configuration,
    KernelParameters&&...     parameters)
{
    // It is assumed arguments were already been validated

    // Note: Unfortunately, even though CUDA should be aware of which context a stream belongs to,
    // and not have trouble enqueueing into a stream in another context - it balks at doing so under
    // certain conditions, so we must place ourselves in the stream's context.
    CAW_SET_SCOPE_CONTEXT(stream.context_handle());
    detail_::enqueue_raw_kernel_launch_in_current_context<RawKernelFunction, KernelParameters...>(
        kernel_function, stream.device_id(), stream.context_handle(), stream.handle(), launch_configuration,
        ::std::forward<KernelParameters>(parameters)...);
}

template<typename Kernel, typename... KernelParameters>
void enqueue_launch(
    bool_constant<true>,  // a kernel wrapped in a kernel_t (sub)class
    bool_constant<false>, // Not a library kernel
    Kernel&&                kernel,
    const stream_t&         stream,
    launch_configuration_t  launch_configuration,
    KernelParameters&&...   parameters)
{
    // It is assumed arguments were already been validated - except for:
#ifndef NDEBUG
    if (kernel.context() != stream.context()) {
        throw ::std::invalid_argument{"Attempt to launch " + kernel::detail_::identify(kernel)
            + " on " + stream::detail_::identify(stream) + ": Different contexts"};
    }
    detail_::validate_compatibility(kernel, launch_configuration);
#endif // #ifndef NDEBUG

    enqueue_launch_helper<typename ::std::decay<Kernel>::type, KernelParameters...>{}(
        ::std::forward<Kernel>(kernel), stream, launch_configuration,
        ::std::forward<KernelParameters>(parameters)...);
}

#if CUDA_VERSION >= 12000
template<typename Kernel, typename... KernelParameters>
void enqueue_launch(
    bool_constant<false>, // Not a wrapped contextual kernel,
    bool_constant<true>,  // but a library kernel
    Kernel&&                kernel,
    const stream_t&         stream,
    launch_configuration_t  launch_configuration,
    KernelParameters&&...   parameters)
{
    // Launch configuration is assumed to have been validated separately
    // from the kernel, and their compatibility will be validated further
    // inside, against the contextualized kernel

    kernel_t contextualized = cuda::contextualize(kernel, stream.context());
    enqueue_launch_helper<kernel_t, KernelParameters...> {}(
        contextualized, stream, launch_configuration,
        ::std::forward<KernelParameters>(parameters)...);
}
#endif // CUDA_VERSION >= 12000

} // namespace detail_

template<typename Kernel, typename... KernelParameters>
void launch(
    Kernel&&                kernel,
    launch_configuration_t  launch_configuration,
    KernelParameters&&...   parameters)
{
    // Argument validation will occur within call to enqueue_launch

    auto primary_context = detail_::get_implicit_primary_context(::std::forward<Kernel>(kernel));
    auto stream = primary_context.default_stream();

    // Note: If Kernel is a kernel_t, and its associated device is different
    // than the current device, the next call will fail:

    enqueue_launch(kernel, stream, launch_configuration, ::std::forward<KernelParameters>(parameters)...);
}

template <typename SpanOfConstVoidPtrLike>
inline void launch_type_erased(
    const kernel_t&         kernel,
    const stream_t&         stream,
    launch_configuration_t  launch_configuration,
    SpanOfConstVoidPtrLike  marshalled_arguments)
{
    // Note: We assume that kernel, stream and launch_configuration have already been validated.
    static_assert(
        ::std::is_same<typename SpanOfConstVoidPtrLike::value_type, void*>::value or
        ::std::is_same<typename SpanOfConstVoidPtrLike::value_type, const void*>::value,
        "The element type of the marshalled arguments container type must be either void* or const void*");
#ifndef NDEBUG
    if (kernel.context() != stream.context()) {
        throw ::std::invalid_argument{"Attempt to launch " + kernel::detail_::identify(kernel)
            + " on " + stream::detail_::identify(stream) + ": Different contexts"};
    }
    detail_::validate_compatibility(kernel, launch_configuration);
    detail_::validate(launch_configuration);
    if (*(marshalled_arguments.end() - 1) != nullptr) {
        throw ::std::invalid_argument("marshalled arguments for a kernel launch must end with a nullptr element");
    }
#endif
    CAW_SET_SCOPE_CONTEXT(stream.context_handle());
    return detail_::enqueue_kernel_launch_by_handle_in_current_context(
        kernel.handle(),
        stream.device_id(),
        stream.context_handle(),
        stream.handle(),
        launch_configuration,
        static_cast<const void**>(marshalled_arguments.data()));
}

#if CUDA_VERSION >= 12000
template <typename SpanOfConstVoidPtrLike>
void launch_type_erased(
    const library::kernel_t&  kernel,
    const stream_t&           stream,
    launch_configuration_t    launch_configuration,
    SpanOfConstVoidPtrLike    marshalled_arguments)
{
    // Argument validation will occur inside the call to launch_type_erased
    auto contextualized = contextualize(kernel, stream.context());
    launch_type_erased(contextualized, stream, launch_configuration, marshalled_arguments);
}
#endif // CUDA_VERSION >= 12000

#if ! CAW_CAN_GET_APRIORI_KERNEL_HANDLE

#if defined(__CUDACC__)

// Unfortunately, the CUDA runtime API does not allow for computation of the grid parameters for maximum occupancy
// from code compiled with a host-side-only compiler! See cuda_runtime.h for details

#if CUDA_VERSION >= 10000
namespace detail_ {

template <typename UnaryFunction>
inline grid::composite_dimensions_t min_grid_params_for_max_occupancy(
    const void *             ptr,
    device::id_t             device_id,
    UnaryFunction            block_size_to_dynamic_shared_mem_size,
    grid::block_dimension_t  block_size_limit,
    bool                     disable_caching_override)
{
    int min_grid_size_in_blocks { 0 };
    int block_size { 0 };
        // Note: only initializing the values her because of a
        // spurious (?) compiler warning about potential uninitialized use.
    auto result = cudaOccupancyMaxPotentialBlockSizeVariableSMemWithFlags(
        &min_grid_size_in_blocks, &block_size,
        ptr,
        block_size_to_dynamic_shared_mem_size,
        static_cast<int>(block_size_limit),
        disable_caching_override ? cudaOccupancyDisableCachingOverride : cudaOccupancyDefault
    );
    throw_if_error_lazy(result,
        "Failed obtaining parameters for a minimum-size grid for kernel " + detail_::ptr_as_hex(ptr) +
            " on device " + ::std::to_string(device_id) + ".");
    return { (grid::dimension_t) min_grid_size_in_blocks, (grid::block_dimension_t) block_size };
}

inline grid::composite_dimensions_t min_grid_params_for_max_occupancy(
    const void *             ptr,
    device::id_t             device_id,
    memory::shared::size_t   dynamic_shared_mem_size,
    grid::block_dimension_t  block_size_limit,
    bool                     disable_caching_override)
{
    auto always_need_same_shared_mem_size =
        [dynamic_shared_mem_size](::size_t) { return dynamic_shared_mem_size; };
    return min_grid_params_for_max_occupancy(
        ptr, device_id, always_need_same_shared_mem_size, block_size_limit, disable_caching_override);
}

} // namespace detail_

inline grid::composite_dimensions_t min_grid_params_for_max_occupancy(
    const kernel::apriori_compiled_t&  kernel,
    memory::shared::size_t            dynamic_shared_memory_size,
    grid::block_dimension_t           block_size_limit,
    bool                              disable_caching_override)
{
    return detail_::min_grid_params_for_max_occupancy(
        kernel.ptr(), kernel.device().id(), dynamic_shared_memory_size, block_size_limit, disable_caching_override);
}

template <typename UnaryFunction>
grid::composite_dimensions_t min_grid_params_for_max_occupancy(
    const kernel::apriori_compiled_t&  kernel,
    UnaryFunction                     block_size_to_dynamic_shared_mem_size,
    grid::block_dimension_t           block_size_limit,
    bool                              disable_caching_override)
{
    return detail_::min_grid_params_for_max_occupancy(
        kernel.ptr(), kernel.device_id(), block_size_to_dynamic_shared_mem_size, block_size_limit, disable_caching_override);
}
#endif // CUDA_VERSION >= 10000

#endif // defined(__CUDACC__)
#endif // ! CAW_CAN_GET_APRIORI_KERNEL_HANDLE

} // namespace cuda

#endif // MULTI_WRAPPER_IMPLS_LAUNCH_HPP_