/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.2/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.2/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.2/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/tensor_description/cluster_descriptor.hpp"
 #include "ck/utility/get_shift.hpp"
 #include "ck/utility/reduction_functions_accumulate.hpp"
  
 namespace ck {
  
 // clang-format off
 // Assume:
 //  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
 //  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
 //  3) in_out_value is the input data in vgpr from each thread
 //  4) in_out_value is the over-written reduced output in vgpr for each thread
 // clang-format on
 template <typename AccDataType,
           index_t BlockSize,
           typename ThreadClusterLengths_M_K,
           typename ThreadClusterArrangeOrder,
           typename OpReduce,
           bool PropagateNan,
           typename Accumulation =
               detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
 struct PartitionedBlockwiseReduction
 {
     static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
                   "The product of cluster lengths should be same as BlockSize!");
  
     static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
     static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
  
     static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
  
     static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
         make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
  
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
  
     template <typename BufferType>
     __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
     {
         static_assert(is_same<typename BufferType::type, AccDataType>{},
                       "Buffer data type should be consistent as AccDataType!");
  
         constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
  
         const auto thread_cluster_idx =
             thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
  
         const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
         const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
  
         work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
  
         __syncthreads();
  
         static_for<0, cluster_len_shift, 1>{}([&](auto I) {
             constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
  
             if(thread_k_cluster_id < indOffset)
             {
                 index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
                 index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
                                                                      make_tuple(0, indOffset));
  
                 AccDataType opData1 = work_buffer[offset1];
                 AccDataType opData2 = work_buffer[offset2];
                 Accumulation::Calculate(opData1, opData2);
                 work_buffer(offset1) = opData1;
             }
  
             __syncthreads();
         });
  
         index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
  
         in_out_value = work_buffer[offset];
     };
 };
  
 // clang-format off
 // Assume:
 //  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
 //  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
 //  3) in_out_value is the input data in vgpr from each thread
 //  4) in_out_value is the over-written reduced output in vgpr for each thread
 // clang-format on
 template <typename AccDataType,
           index_t BlockSize,
           typename ThreadClusterLengths_M_K,
           typename ThreadClusterDesc,
           typename OpReduce,
           bool PropagateNan,
           typename Accumulation =
               detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
 struct PartitionedBlockwiseReduction_v2
 {
     static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
                   "The product of cluster lengths should be same as BlockSize!");
  
     static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
     static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
  
     static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
  
     static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
         make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
  
     static constexpr auto thread_cluster_desc = ThreadClusterDesc{};
  
     template <typename BufferType>
     __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
     {
         static_assert(is_same<typename BufferType::type, AccDataType>{},
                       "Buffer data type should be consistent as AccDataType!");
  
         constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
  
         const auto thread_cluster_idx =
             thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
  
         const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
         const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
  
         work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
  
         __syncthreads();
  
         static_for<0, cluster_len_shift, 1>{}([&](auto I) {
             constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
  
             if(thread_k_cluster_id < indOffset)
             {
                 index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
                 index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
                                                                      make_tuple(0, indOffset));
  
                 AccDataType opData1 = work_buffer[offset1];
                 AccDataType opData2 = work_buffer[offset2];
                 Accumulation::Calculate(opData1, opData2);
                 work_buffer(offset1) = opData1;
             }
  
             __syncthreads();
         });
  
         index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
  
         in_out_value = work_buffer[offset];
     };
 };
  
 // clang-format off
 // Assume:
 //  1) work_val_buffer/work_idx_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
 //  2) work_val_buffer/work_idx_buffer has AccDataType/IndexDataType elements, and space size is no less than BlockSize
 //  3) in_out_value/in_out_index is the input data in vgpr from each thread
 //  4) in_out_value/in_out_index is the over-written reduced output in vgpr for each thread
 // clang-format on
 template <
     typename AccDataType,
     typename IndexDataType,
     index_t BlockSize,
     typename ThreadClusterLengths_M_K,
     typename ThreadClusterArrangeOrder,
     typename OpReduce,
     bool PropagateNan,
     typename Accumulation =
         detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
 struct PartitionedBlockwiseReductionWithIndex
 {
     static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
                   "The product of cluster lengths should be same as BlockSize!");
  
     static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
     static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
  
     static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
  
     static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
         make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
  
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
  
     // This interface accumulates on both data values and indices
     template <typename BufferType, typename IdxBufferType>
     __device__ static void Reduce(BufferType& work_val_buffer,
                                   IdxBufferType& work_idx_buffer,
                                   AccDataType& in_out_value,
                                   IndexDataType& in_out_index)
     {
         static_assert(is_same<typename BufferType::type, AccDataType>{},
                       "Buffer data type should be consistent as AccDataType!");
         static_assert(is_same<typename IdxBufferType::type, IndexDataType>{},
                       "Buffer data type should be consistent as IndexDataType!");
  
         constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
  
         const auto thread_cluster_idx =
             thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
  
         const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
         const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
  
         work_val_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
         work_idx_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_index;
  
         __syncthreads();
  
         static_for<0, cluster_len_shift, 1>{}([&](auto I) {
             constexpr index_t indOffset = 1 << I();
  
             if(thread_k_cluster_id % (indOffset * 2) == 0)
             {
                 index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
                 index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
                                                                      make_tuple(0, indOffset));
  
                 AccDataType opData1      = work_val_buffer[offset1];
                 AccDataType opData2      = work_val_buffer[offset2];
                 IndexDataType currIndex1 = work_idx_buffer[offset1];
                 IndexDataType currIndex2 = work_idx_buffer[offset2];
  
                 Accumulation::Calculate(opData1, opData2, currIndex1, currIndex2);
                 work_val_buffer(offset1) = opData1;
                 work_idx_buffer(offset1) = currIndex1;
             }
  
             __syncthreads();
         });
  
         index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
  
         in_out_value = work_val_buffer[offset];
         in_out_index = work_idx_buffer[offset];
     };
 };
  
 } // namespace ck