algos.h 7.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/**
 * \file dnn/src/fallback/convolution/algos.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */

#pragma once

#include "src/fallback/conv_bias/algos.h"
#include "src/fallback/convolution/opr_impl.h"
#include "src/naive/convolution/helper.h"

namespace megdnn {
namespace fallback {

template <typename ST, typename DT, typename CT>
void kern_naive_forward(const ConvolutionImpl::NCBKernParam& p,
23 24 25
                        const ConvolutionImpl::NCBKernIndex& ncb_index) {
    size_t batch_id = ncb_index.ndrange_id[1];
    size_t group_id = ncb_index.ndrange_id[0];
26 27
    auto IC = p.filter_meta.icpg, IH = p.isz[0], IW = p.isz[1],
         OC = p.filter_meta.ocpg, OH = p.osz[0], OW = p.osz[1];
28 29 30 31 32
    ptrdiff_t fstrd = p.filter_meta.icpg * p.filter_meta.ocpg *
                      p.filter_meta.spatial[0] * p.filter_meta.spatial[1] *
                      p.filter_type.size();
    ptrdiff_t istrd = p.filter_meta.icpg * p.src_type.size();
    ptrdiff_t ostrd = p.filter_meta.ocpg * p.dst_type.size();
33 34 35 36 37
    TensorND src, dst;

    src.layout.dtype = p.src_type;
    dst.layout.dtype = p.dst_type;
    if (p.filter_meta.format == param::Convolution::Format::NCHW) {
38 39 40 41
       istrd *= p.isz[0] * p.isz[1];
       ostrd *= p.osz[0] * p.osz[1];
       src.layout.init_contiguous_stride({1, IC, IH, IW});
       dst.layout.init_contiguous_stride({1, OC, OH, OW});
42 43 44 45 46 47 48 49 50
    } else {
        // Must be NHWC
        megdnn_assert(
                p.filter_meta.format == param::Convolution::Format::NHWC,
                "AlgoNaive only support NCHW and NHWC, not support format %d",
                static_cast<int>(p.filter_meta.format));
        src.layout.init_contiguous_stride({1, IH, IW, IC});
        dst.layout.init_contiguous_stride({1, OH, OW, OC});
    }
51 52 53 54 55 56 57 58
    src.raw_ptr = reinterpret_cast<void*>(
            reinterpret_cast<uintptr_t>(p.src_ptr) +
            batch_id * p.inp_bs * p.src_type.size() + group_id * istrd);
    dst.raw_ptr = reinterpret_cast<void*>(
            reinterpret_cast<uintptr_t>(p.dst_ptr) +
            batch_id * p.out_bs * p.dst_type.size() + group_id * ostrd);
    ST* filter = reinterpret_cast<ST*>(
            reinterpret_cast<uintptr_t>(p.filter_ptr) + group_id * fstrd);
59 60
    std::copy(p.inp_s, p.inp_s + 4, src.layout.stride);
    std::copy(p.out_s, p.out_s + 4, dst.layout.stride);
61
    naive::convolution::forward<ST, ST, DT, CT>(src, filter, dst,
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
                                                p.filter_meta);
}

template <typename ftype, typename dtype, typename gtype>
void kern_naive(const ConvolutionBackwardDataImpl::NCBKernParam& p) {
    TensorND diff(const_cast<void*>(p.diff_ptr), p.diff_layout),
            filter(const_cast<void*>(p.filter_ptr), p.filter_layout),
            grad(p.grad_ptr, p.grad_layout);
    naive::convolution::backward_data<ftype, dtype, gtype>(filter, diff, grad,
                                                           p.filter_meta);
}

class ConvolutionImpl::AlgoFallback final : public AlgoBase {
public:
    bool is_reproducible() const override { return true; }
    const char* name() const override { return "FALLBACK_ALGO"; }
    bool usable(ConvolutionImpl* opr, const NCBKernSizeParam& param,
                AlgoSelectionStrategy algo_selection_strategy) const override;

    size_t get_workspace(ConvolutionImpl* opr,
                         const NCBKernSizeParam& param) const override;

    SmallVector<NCBKern> dispatch_kern(
            ConvolutionImpl* /*opr*/,
            const NCBKernSizeParam& /*param*/) const override;
};

class ConvolutionImpl::AlgoNaive final : public AlgoBase {
public:
    bool is_reproducible() const override { return true; }
    const char* name() const override { return "NAIVE_ALGO"; }
    bool usable(ConvolutionImpl* /*opr*/, const NCBKernSizeParam& /*param*/,
                AlgoSelectionStrategy algo_selection_strategy) const override;

    size_t get_workspace(ConvolutionImpl*,
                         const NCBKernSizeParam&) const override {
        return 0;
    };

    SmallVector<NCBKern> dispatch_kern(
            ConvolutionImpl* /*opr*/,
            const NCBKernSizeParam& /*param*/) const override;
};

class ConvolutionImpl::AlgoDefault final : public AlgoBase {
    static ConvBiasImpl::NCBKernSizeParam init_convbias_opr_and_param(
            ConvBiasImpl* conv_bias_opr, const NCBKernSizeParam& param);
    WorkspaceBundle get_bundle(const NCBKernSizeParam& param) const;
    static SmallVector<NCBKern> get_kimpl(ConvBiasImpl* conv_bias_opr,
                                          ConvBiasImpl::AlgoBase* algo,
                                          const NCBKernSizeParam& param);

public:
    AlgoDefault(fallback::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase*);
    bool is_reproducible() const override { return true; }
    const char* name() const override { return m_name.c_str(); }
    bool usable(ConvolutionImpl* opr, const NCBKernSizeParam& param,
                AlgoSelectionStrategy algo_selection_strategy) const override;

    size_t get_workspace(ConvolutionImpl* opr,
                         const NCBKernSizeParam& param) const override;

    SmallVector<NCBKern> dispatch_kern(
            ConvolutionImpl* /*opr*/,
            const NCBKernSizeParam& param) const override {
        return get_kimpl(m_conv_bias_opr, m_algorithm, param);
    }

    void* type() const override { return sm_fallback_conv_algo_type; }

    //! select matmul to the highest preference
    bool is_preferred(ConvolutionImpl* opr,
                      const NCBKernSizeParam& param) const override;

private:
    std::string m_name;
    fallback::ConvBiasImpl* m_conv_bias_opr;
    ConvBiasImpl::AlgoBase* m_algorithm;
};

class ConvolutionBackwardDataImpl::AlgoDirect final : public AlgoBase {
public:
    bool is_reproducible() const override { return true; }
    const char* name() const override { return "DeconvDirect"; }
    bool usable(ConvolutionBackwardDataImpl* opr,
                const NCBKernSizeParam& param) const override;
    size_t get_workspace(ConvolutionBackwardDataImpl*,
                         const NCBKernSizeParam& param) const override;
    ncb_kern_t dispatch_kern(ConvolutionBackwardDataImpl*,
                             const NCBKernSizeParam&) const override;
    void* type() const override { return sm_fallback_deconv_algo_type; }
};

class ConvolutionBackwardDataImpl::AlgoMatrixMul final : public AlgoBase {
public:
    bool is_reproducible() const override { return true; }
    const char* name() const override { return "DeconvMatmul"; }
    bool usable(ConvolutionBackwardDataImpl* opr,
                const NCBKernSizeParam& param) const override;
    size_t get_workspace(ConvolutionBackwardDataImpl*,
                         const NCBKernSizeParam& param) const override;
    ncb_kern_t dispatch_kern(ConvolutionBackwardDataImpl*,
                             const NCBKernSizeParam&) const override;
    void* type() const override { return sm_fallback_deconv_algo_type; }
};

}  // namespace fallback
}  // namespace megdnn

// vim: syntax=cpp.doxygen