direct_convolution.cuh

#pragma once
#include "constant_tensor_descriptor.cuh"
#include "blockwise_tensor_op.cuh"
#include "threadwise_tensor_op.cuh"

template <class TFloat, class InDesc, class WeiDesc, class OutDesc>
__device__ void threadwise_direct_convolution(InDesc,
                                              TFloat* const __restrict__ p_in,
                                              WeiDesc,
                                              TFloat* const __restrict__ p_wei,
                                              OutDesc,
                                              TFloat* __restrict__ p_out)
{
    constexpr auto I0 = Index<0>{};
    constexpr auto I1 = Index<1>{};
    constexpr auto I2 = Index<2>{};
    constexpr auto I3 = Index<3>{};

    constexpr auto in_desc  = InDesc{};
    constexpr auto wei_desc = WeiDesc{};
    constexpr auto out_desc = OutDesc{};

#if 0
    if(threadIdx.x == 0)
    {
        print_ConstantTensorDescriptor(in_desc);
        print_ConstantTensorDescriptor(wei_desc);
        print_ConstantTensorDescriptor(out_desc);
    }
#endif

    for(unsigned n = 0; n < out_desc.GetLength(I0); ++n)
    {
        for(unsigned k = 0; k < out_desc.GetLength(I1); ++k)
        {
            for(unsigned ho = 0; ho < out_desc.GetLength(I2); ++ho)
            {
                for(unsigned wo = 0; wo < out_desc.GetLength(I3); ++wo)
                {
                    for(unsigned c = 0; c < wei_desc.GetLength(I1); ++c)
                    {
                        for(unsigned s = 0; s < wei_desc.GetLength(I2); ++s)
                        {
                            for(unsigned r = 0; r < wei_desc.GetLength(I3); ++r)
                            {
                                const unsigned hi = ho + s;
                                const unsigned wi = wo + r;

                                const unsigned in_index =
                                    in_desc.GetStride(I0) * n + in_desc.GetStride(I1) * c +
                                    in_desc.GetStride(I2) * hi + in_desc.GetStride(I3) * wi;

                                const unsigned wei_index =
                                    wei_desc.GetStride(I0) * k + wei_desc.GetStride(I1) * c +
                                    wei_desc.GetStride(I2) * s + in_desc.GetStride(I3) * r;

                                const unsigned out_index =
                                    out_desc.GetStride(I0) * n + out_desc.GetStride(I1) * k +
                                    out_desc.GetStride(I2) * ho + out_desc.GetStride(I3) * wo;

                                p_out[out_index] += p_wei[wei_index] * p_in[in_index];

#if 0
                                if(threadIdx.x == 0)
                                {
                                    printf("threadwise_direct_convolution: 1: \t"
                                           "threadIdx.x %u\t"
                                           "out_index %u, p_out[out_index] %f, \t"
                                           "wei_index %u, p_wei[wei_index] %f, \t"
                                           "in_index %u, p_in[in_index] %f\n",
                                           threadIdx.x,
                                           out_index,
                                           p_out[out_index],
                                           wei_index,
                                           p_wei[wei_index],
                                           in_index,
                                           p_in[in_index]);
                                }
#endif
                            }
                        }
                    }
                }
            }
        }
    }
}

template <class TFloat,
          class InDesc,
          class WeiDesc,
          class OutDesc,
          unsigned OutTileSizeH,
          unsigned OutTileSizeW,
          unsigned BlockSize>
__device__ void blockwise_convolution(InDesc,
                                      TFloat* const __restrict__ p_in,
                                      WeiDesc,
                                      TFloat* const __restrict__ p_wei,
                                      OutDesc,
                                      TFloat* __restrict__ p_out)
{
    constexpr auto I0 = Index<0>{};
    constexpr auto I1 = Index<1>{};
    constexpr auto I2 = Index<2>{};
    constexpr auto I3 = Index<3>{};

    constexpr auto in_desc  = InDesc{};
    constexpr auto wei_desc = WeiDesc{};
    constexpr auto out_desc = OutDesc{};

    constexpr unsigned S = wei_desc.GetLength(I2);
    constexpr unsigned R = wei_desc.GetLength(I3);

    constexpr unsigned NPerBlock = out_desc.GetLength(I0);
    constexpr unsigned KPerBlock = out_desc.GetLength(I1);
    constexpr unsigned YPerBlock = (out_desc.GetLength(I2) + OutTileSizeH - 1) / OutTileSizeH;
    constexpr unsigned XPerBlock = (out_desc.GetLength(I3) + OutTileSizeW - 1) / OutTileSizeW;

    constexpr unsigned CPerBlockLoop = in_desc.GetLength(I1);

    constexpr unsigned InTileSizeH = OutTileSizeH + S - 1;
    constexpr unsigned InTileSizeW = OutTileSizeW + R - 1;

#if 0
    if(threadIdx.x == 0)
    {
        print_ConstantTensorDescriptor(in_desc);
        print_ConstantTensorDescriptor(wei_desc);
        print_ConstantTensorDescriptor(out_desc);
    }
#endif

    constexpr auto in_thread_src_desc = make_ConstantTensorDescriptor(
        Sequence<1, CPerBlockLoop, OutTileSizeH + S - 1, OutTileSizeW + R - 1>{},
        in_desc.GetStrides());

    constexpr auto wei_thread_src_desc =
        make_ConstantTensorDescriptor(Sequence<1, CPerBlockLoop, S, R>{}, wei_desc.GetStrides());

    constexpr auto out_thread_src_desc = make_ConstantTensorDescriptor(
        Sequence<1, 1, OutTileSizeH, OutTileSizeW>{}, out_desc.GetStrides());

    constexpr auto in_thread_dst_desc =
        make_ConstantTensorDescriptor(in_thread_src_desc.GetLengths());

    constexpr auto wei_thread_dst_desc =
        make_ConstantTensorDescriptor(wei_thread_src_desc.GetLengths());

    constexpr auto out_thread_dst_desc =
        make_ConstantTensorDescriptor(out_thread_src_desc.GetLengths());

    const unsigned thread_id = threadIdx.x;

    for(unsigned thread_work_id = thread_id;
        thread_work_id < NPerBlock * KPerBlock * YPerBlock * XPerBlock;
        thread_work_id += BlockSize)
    {
        unsigned itmp             = thread_work_id;
        unsigned n_thread_work_id = itmp / (KPerBlock * YPerBlock * XPerBlock);
        itmp -= n_thread_work_id * (KPerBlock * YPerBlock * XPerBlock);
        unsigned k_thread_work_id = itmp / (YPerBlock * XPerBlock);
        itmp -= k_thread_work_id * (YPerBlock * XPerBlock);
        unsigned y_thread_work_id = itmp / XPerBlock;
        unsigned x_thread_work_id = itmp - y_thread_work_id * XPerBlock;

        unsigned n_thread_work_begin  = n_thread_work_id * 1;
        unsigned k_thread_work_begin  = k_thread_work_id * 1;
        unsigned ho_thread_work_begin = y_thread_work_id * OutTileSizeH;
        unsigned wo_thread_work_begin = x_thread_work_id * OutTileSizeW;

        unsigned hi_thread_work_begin = ho_thread_work_begin; // minus padding
        unsigned wi_thread_work_begin = wo_thread_work_begin; // minus padding

        TFloat p_in_thread[1 * CPerBlockLoop * InTileSizeH * InTileSizeW];
        TFloat p_wei_thread[1 * CPerBlockLoop * S * R];
        TFloat p_out_thread[1 * 1 * OutTileSizeH * OutTileSizeW];

        auto f_copy = [](const TFloat& src, TFloat& dst) { dst = src; };

        // copy input tensor into register
        threadwise_4d_tensor_op<TFloat,
                                decltype(in_thread_src_desc),
                                decltype(in_thread_dst_desc),
                                decltype(f_copy)>(
            in_thread_src_desc,
            p_in + in_desc.Get1dIndex(
                       n_thread_work_begin, 0, hi_thread_work_begin, wi_thread_work_begin),
            in_thread_dst_desc,
            p_in_thread,
            f_copy);

        // copy weight tensor into register
        threadwise_4d_tensor_op<TFloat,
                                decltype(wei_thread_src_desc),
                                decltype(wei_thread_dst_desc),
                                decltype(f_copy)>(
            wei_thread_src_desc,
            p_wei + wei_desc.Get1dIndex(k_thread_work_begin, 0, 0, 0),
            wei_thread_dst_desc,
            p_wei_thread,
            f_copy);

        // copy output tensor into register
        threadwise_4d_tensor_op<TFloat,
                                decltype(out_thread_src_desc),
                                decltype(out_thread_dst_desc),
                                decltype(f_copy)>(out_thread_src_desc,
                                                  p_out + out_desc.Get1dIndex(n_thread_work_begin,
                                                                              k_thread_work_begin,
                                                                              ho_thread_work_begin,
                                                                              wo_thread_work_begin),
                                                  out_thread_dst_desc,
                                                  p_out_thread,
                                                  f_copy);

        // threadwise convolution
        threadwise_direct_convolution<TFloat,
                                      decltype(in_thread_dst_desc),
                                      decltype(wei_thread_dst_desc),
                                      decltype(out_thread_dst_desc)>(in_thread_dst_desc,
                                                                     p_in_thread,
                                                                     wei_thread_dst_desc,
                                                                     p_wei_thread,
                                                                     out_thread_dst_desc,
                                                                     p_out_thread);

        // accumulate output tensor into device mem
        threadwise_4d_tensor_op<TFloat,
                                decltype(out_thread_dst_desc),
                                decltype(out_thread_src_desc),
                                decltype(f_copy)>(out_thread_dst_desc,
                                                  p_out_thread,
                                                  out_thread_src_desc,
                                                  p_out + out_desc.Get1dIndex(n_thread_work_begin,
                                                                              k_thread_work_begin,
                                                                              ho_thread_work_begin,
                                                                              wo_thread_work_begin),
                                                  f_copy);
    }
}

template <class TFloat,
          class InDesc,
          class WeiDesc,
          class OutDesc,
          unsigned NPerBlock,
          unsigned KPerBlock,
          unsigned CPerBlockLoop,
          unsigned OutTileSizeH,
          unsigned OutTileSizeW,
          unsigned YPerBlock,
          unsigned XPerBlock,
          unsigned NBlockCopyLen0,
          unsigned NBlockCopyLen1,
          unsigned NBlockCopyLen2,
          unsigned NBlockCopyLen3,
          unsigned BlockSize,
          unsigned GridSize>
__global__ void gridwise_convolution(InDesc,
                                     TFloat* const __restrict__ p_in,
                                     WeiDesc,
                                     TFloat* const __restrict__ p_wei,
                                     OutDesc,
                                     TFloat* __restrict__ p_out)
{
    constexpr auto I0 = Index<0>{};
    constexpr auto I1 = Index<1>{};
    constexpr auto I2 = Index<2>{};
    constexpr auto I3 = Index<3>{};

    constexpr auto in_desc  = InDesc{};
    constexpr auto wei_desc = WeiDesc{};
    constexpr auto out_desc = OutDesc{};

    constexpr unsigned S = wei_desc.GetLength(I2);
    constexpr unsigned R = wei_desc.GetLength(I3);

    constexpr unsigned HoPerBlock = OutTileSizeH * YPerBlock;
    constexpr unsigned WoPerBlock = OutTileSizeW * XPerBlock;

    constexpr unsigned HiPerBlock = YPerBlock * OutTileSizeH + S - 1;
    constexpr unsigned WiPerBlock = XPerBlock * OutTileSizeW + R - 1;

    constexpr unsigned NBlockWork = (out_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
    constexpr unsigned KBlockWork = (out_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
    constexpr unsigned YBlockWork = (out_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
    constexpr unsigned XBlockWork = (out_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;

    constexpr auto in_block_glb_desc = make_ConstantTensorDescriptor(
        Sequence<NPerBlock, CPerBlockLoop, HiPerBlock, WiPerBlock>{}, in_desc.GetStrides());

    constexpr auto wei_block_glb_desc = make_ConstantTensorDescriptor(
        Sequence<KPerBlock, CPerBlockLoop, S, R>{}, wei_desc.GetStrides());

    constexpr auto out_block_glb_desc = make_ConstantTensorDescriptor(
        Sequence<NPerBlock, KPerBlock, HoPerBlock, WoPerBlock>{}, out_desc.GetStrides());

    constexpr auto in_block_lds_desc =
        make_ConstantTensorDescriptor(in_block_glb_desc.GetLengths());
    constexpr auto wei_block_lds_desc =
        make_ConstantTensorDescriptor(wei_block_glb_desc.GetLengths());
    constexpr auto out_block_lds_desc =
        make_ConstantTensorDescriptor(out_block_glb_desc.GetLengths());

    constexpr unsigned in_block_size  = in_block_lds_desc.GetElementSize();
    constexpr unsigned wei_block_size = wei_block_lds_desc.GetElementSize();
    constexpr unsigned out_block_size = out_block_lds_desc.GetElementSize();

    __shared__ TFloat p_in_block[in_block_size];
    __shared__ TFloat p_wei_block[wei_block_size];
    __shared__ TFloat p_out_block[out_block_size];

    const unsigned block_id = blockIdx.x;

    unsigned itmp            = block_id;
    unsigned n_block_work_id = itmp / (KBlockWork * YBlockWork * XBlockWork);
    itmp -= n_block_work_id * (KBlockWork * YBlockWork * XBlockWork);
    unsigned k_block_work_id = itmp / (YBlockWork * XBlockWork);
    itmp -= k_block_work_id * (YBlockWork * XBlockWork);
    unsigned y_block_work_id = itmp / XBlockWork;
    unsigned x_block_work_id = itmp - y_block_work_id * XBlockWork;

    unsigned n_block_work_begin = n_block_work_id * NPerBlock;
    unsigned k_block_work_begin = k_block_work_id * KPerBlock;
    unsigned y_block_work_begin = y_block_work_id * YPerBlock;
    unsigned x_block_work_begin = x_block_work_id * XPerBlock;

    unsigned ho_block_work_begin = y_block_work_begin * OutTileSizeH;
    unsigned wo_block_work_begin = x_block_work_begin * OutTileSizeW;

    unsigned hi_block_work_begin = ho_block_work_begin; // minus padding
    unsigned wi_block_work_begin = wo_block_work_begin; // minus padding

#if 0
    if(threadIdx.x == 0)
    {
        print_ConstantTensorDescriptor( in_desc, "gridwise_convolution:  in_desc: ");
        print_ConstantTensorDescriptor(wei_desc, "gridwise_convolution: wei_desc: ");
        print_ConstantTensorDescriptor(out_desc, "gridwise_convolution: out_desc: ");
        print_ConstantTensorDescriptor( in_block_glb_desc, "gridwise_convolution:  in_block_glb_desc: ");
        print_ConstantTensorDescriptor(wei_block_glb_desc, "gridwise_convolution: wei_block_glb_desc: ");
        print_ConstantTensorDescriptor(out_block_glb_desc, "gridwise_convolution: out_block_glb_desc: ");
        print_ConstantTensorDescriptor( in_block_lds_desc, "gridwise_convolution:  in_block_lds_desc: ");
        print_ConstantTensorDescriptor(wei_block_lds_desc, "gridwise_convolution: wei_block_lds_desc: ");
        print_ConstantTensorDescriptor(out_block_lds_desc, "gridwise_convolution: out_block_lds_desc: ");

        printf("NBlockWork %u, KBlockWork %u, YBlockWork %u, XBlockWork %u \t"
               "block_id %u, n_block_work_id %u, k_block_work_id %u, y_block_work_id %u, "
               "x_block_work_id %u\n",
               NBlockWork,
               KBlockWork,
               YBlockWork,
               XBlockWork,
               block_id,
               n_block_work_id,
               k_block_work_id,
               y_block_work_id,
               x_block_work_id);
    }
#endif

    for(unsigned c_block_work_begin = 0; c_block_work_begin < in_desc.GetLength(I1);
        c_block_work_begin += CPerBlockLoop)
    {
        auto f_copy = [](const TFloat& src, TFloat& dst) { dst = src; };

        // copy input tensor to LDS
        blockwise_4d_tensor_op<TFloat,
                               decltype(in_block_glb_desc),
                               decltype(in_block_lds_desc),
                               NBlockCopyLen0,
                               NBlockCopyLen1,
                               NBlockCopyLen2,
                               NBlockCopyLen3,
                               decltype(f_copy),
                               BlockSize>(in_block_glb_desc,
                                          p_in + in_block_glb_desc.Get1dIndex(n_block_work_begin,
                                                                              c_block_work_begin,
                                                                              hi_block_work_begin,
                                                                              wi_block_work_begin),
                                          in_block_lds_desc,
                                          p_in_block,
                                          f_copy);

        // copy weight tensor to LDS
        blockwise_4d_tensor_op<TFloat,
                               decltype(wei_block_glb_desc),
                               decltype(wei_block_lds_desc),
                               NBlockCopyLen0,
                               NBlockCopyLen1,
                               NBlockCopyLen2,
                               NBlockCopyLen3,
                               decltype(f_copy),
                               BlockSize>(
            wei_block_glb_desc,
            p_wei + wei_block_glb_desc.Get1dIndex(k_block_work_begin, c_block_work_begin, 0, 0),
            wei_block_lds_desc,
            p_wei_block,
            f_copy);

        // copy output tensor to LDS
        blockwise_4d_tensor_op<TFloat,
                               decltype(out_block_glb_desc),
                               decltype(out_block_lds_desc),
                               NBlockCopyLen0,
                               NBlockCopyLen1,
                               NBlockCopyLen2,
                               NBlockCopyLen3,
                               decltype(f_copy),
                               BlockSize>(out_block_glb_desc,
                                          p_out +
                                              out_block_glb_desc.Get1dIndex(n_block_work_begin,
                                                                            k_block_work_begin,
                                                                            ho_block_work_begin,
                                                                            wo_block_work_begin),
                                          out_block_lds_desc,
                                          p_out_block,
                                          f_copy);

#if 1
        __syncthreads();
#endif

        // blockwise convolution
        blockwise_convolution<TFloat,
                              decltype(in_block_lds_desc),
                              decltype(wei_block_lds_desc),
                              decltype(out_block_lds_desc),
                              OutTileSizeH,
                              OutTileSizeW,
                              BlockSize>(in_block_lds_desc,
                                         p_in_block,
                                         wei_block_lds_desc,
                                         p_wei_block,
                                         out_block_lds_desc,
                                         p_out_block);

#if 1
        __syncthreads();
#endif

        // accum output tensor from LDS to device mem
        blockwise_4d_tensor_op<TFloat,
                               decltype(out_block_lds_desc),
                               decltype(out_block_glb_desc),
                               NBlockCopyLen0,
                               NBlockCopyLen1,
                               NBlockCopyLen2,
                               NBlockCopyLen3,
                               decltype(f_copy),
                               BlockSize>(out_block_lds_desc,
                                          p_out_block,
                                          out_block_glb_desc,
                                          p_out +
                                              out_block_glb_desc.Get1dIndex(n_block_work_begin,
                                                                            k_block_work_begin,
                                                                            ho_block_work_begin,
                                                                            wo_block_work_begin),
                                          f_copy);
    }
}