/home/runner/work/cccl/cccl/cub/cub/util_vsmem.cuh
File members: /home/runner/work/cccl/cccl/cub/cub/util_vsmem.cuh
/******************************************************************************
* Copyright (c) 2023-24, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the
* names of its contributors may be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
* DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
******************************************************************************/
#pragma once
#include <cub/config.cuh>
#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
# pragma GCC system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
# pragma clang system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
# pragma system_header
#endif // no system header
#include <cub/util_device.cuh>
#include <cub/util_ptx.cuh>
#include <cub/util_type.cuh>
#include <cuda/discard_memory>
#include <cstdint>
CUB_NAMESPACE_BEGIN
#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
namespace detail
{
struct vsmem_t
{
void* gmem_ptr;
};
// The maximum amount of static shared memory available per thread block
// Note that in contrast to dynamic shared memory, static shared memory is still limited to 48 KB
static constexpr std::size_t max_smem_per_block = 48 * 1024;
template <typename AgentT>
class vsmem_helper_impl
{
private:
// Per-block virtual shared memory may be padded to make sure vsmem is an integer multiple of `line_size`
static constexpr std::size_t line_size = 128;
// The amount of shared memory or virtual shared memory required by the algorithm's agent
static constexpr std::size_t required_smem = sizeof(typename AgentT::TempStorage);
// Whether we need to allocate global memory-backed virtual shared memory
static constexpr bool needs_vsmem = required_smem > max_smem_per_block;
// Padding bytes to an integer multiple of `line_size`. Only applies to virtual shared memory
static constexpr std::size_t padding_bytes =
(required_smem % line_size == 0) ? 0 : (line_size - (required_smem % line_size));
public:
// Type alias to be used for static temporary storage declaration within the algorithm's kernel
using static_temp_storage_t = cub::detail::conditional_t<needs_vsmem, cub::NullType, typename AgentT::TempStorage>;
// The amount of global memory-backed virtual shared memory needed, padded to an integer multiple of 128 bytes
static constexpr std::size_t vsmem_per_block = needs_vsmem ? (required_smem + padding_bytes) : 0;
static _CCCL_DEVICE _CCCL_FORCEINLINE typename AgentT::TempStorage&
get_temp_storage(typename AgentT::TempStorage& static_temp_storage, vsmem_t&)
{
return static_temp_storage;
}
static _CCCL_DEVICE _CCCL_FORCEINLINE typename AgentT::TempStorage&
get_temp_storage(typename AgentT::TempStorage& static_temp_storage, vsmem_t&, std::size_t)
{
return static_temp_storage;
}
static _CCCL_DEVICE _CCCL_FORCEINLINE typename AgentT::TempStorage&
get_temp_storage(cub::NullType& static_temp_storage, vsmem_t& vsmem)
{
return *reinterpret_cast<typename AgentT::TempStorage*>(
static_cast<char*>(vsmem.gmem_ptr) + (vsmem_per_block * blockIdx.x));
}
static _CCCL_DEVICE _CCCL_FORCEINLINE typename AgentT::TempStorage&
get_temp_storage(cub::NullType& static_temp_storage, vsmem_t& vsmem, std::size_t linear_block_id)
{
return *reinterpret_cast<typename AgentT::TempStorage*>(
static_cast<char*>(vsmem.gmem_ptr) + (vsmem_per_block * linear_block_id));
}
template <bool needs_vsmem_ = needs_vsmem, typename ::cuda::std::enable_if<!needs_vsmem_, int>::type = 0>
static _CCCL_DEVICE _CCCL_FORCEINLINE bool discard_temp_storage(typename AgentT::TempStorage& temp_storage)
{
return false;
}
template <bool needs_vsmem_ = needs_vsmem, typename ::cuda::std::enable_if<needs_vsmem_, int>::type = 0>
static _CCCL_DEVICE _CCCL_FORCEINLINE bool discard_temp_storage(typename AgentT::TempStorage& temp_storage)
{
// Ensure all threads finished using temporary storage
CTA_SYNC();
const std::size_t linear_tid = threadIdx.x;
const std::size_t block_stride = line_size * blockDim.x;
char* ptr = reinterpret_cast<char*>(&temp_storage);
auto ptr_end = ptr + vsmem_per_block;
// 128 byte-aligned virtual shared memory discard
for (auto thread_ptr = ptr + (linear_tid * line_size); thread_ptr < ptr_end; thread_ptr += block_stride)
{
cuda::discard_memory(thread_ptr, line_size);
}
return true;
}
};
template <class DefaultAgentT, class FallbackAgentT>
constexpr bool use_fallback_agent()
{
return (sizeof(typename DefaultAgentT::TempStorage) > max_smem_per_block)
&& (sizeof(typename FallbackAgentT::TempStorage) <= max_smem_per_block);
}
template <typename DefaultAgentPolicyT,
typename DefaultAgentT,
typename FallbackAgentPolicyT = DefaultAgentPolicyT,
typename FallbackAgentT = DefaultAgentT,
bool UseFallbackPolicy = use_fallback_agent<DefaultAgentT, FallbackAgentT>()>
struct vsmem_helper_with_fallback_impl : public vsmem_helper_impl<DefaultAgentT>
{
using agent_t = DefaultAgentT;
using agent_policy_t = DefaultAgentPolicyT;
};
template <typename DefaultAgentPolicyT, typename DefaultAgentT, typename FallbackAgentPolicyT, typename FallbackAgentT>
struct vsmem_helper_with_fallback_impl<DefaultAgentPolicyT, DefaultAgentT, FallbackAgentPolicyT, FallbackAgentT, true>
: public vsmem_helper_impl<FallbackAgentT>
{
using agent_t = FallbackAgentT;
using agent_policy_t = FallbackAgentPolicyT;
};
template <typename DefaultPolicyT, typename FallbackPolicyT, template <typename...> class AgentT, typename... AgentParamsT>
using vsmem_helper_fallback_policy_t =
vsmem_helper_with_fallback_impl<DefaultPolicyT,
AgentT<DefaultPolicyT, AgentParamsT...>,
FallbackPolicyT,
AgentT<FallbackPolicyT, AgentParamsT...>>;
template <typename DefaultPolicyT, template <typename...> class AgentT, typename... AgentParamsT>
using vsmem_helper_default_fallback_policy_t =
vsmem_helper_fallback_policy_t<DefaultPolicyT, policy_wrapper_t<DefaultPolicyT, 64, 1>, AgentT, AgentParamsT...>;
} // namespace detail
#endif // DOXYGEN_SHOULD_SKIP_THIS
CUB_NAMESPACE_END