gemm_x8s8s32x_convolution.hpp 12.4 KB
Newer Older
Alexey Suhov's avatar
Alexey Suhov committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
/*******************************************************************************
* Copyright 2017-2019 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#ifndef GEMM_X8S8S32X_CONVOLUTION_HPP
#define GEMM_X8S8S32X_CONVOLUTION_HPP

#include "c_types_map.hpp"
#include "memory_tracking.hpp"

#include "cpu_convolution_pd.hpp"
#include "cpu_engine.hpp"
#include "jit_primitive_conf.hpp"
#include "jit_generator.hpp"
#include "gemm_convolution_utils.hpp"

#include "gemm/gemm.hpp"

namespace mkldnn {
namespace impl {
namespace cpu {

template <data_type_t src_type, data_type_t dst_type>
struct _gemm_x8s8s32x_convolution_fwd_t: public cpu_primitive_t {
    struct pd_t: public cpu_convolution_fwd_pd_t {
        pd_t(engine_t *engine, const convolution_desc_t *adesc,
                const primitive_attr_t *attr,
                const typename pd_t::base_class *hint_fwd_pd)
            : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
            , jcp_() {}

        DECLARE_COMMON_PD_T(IGEMM_S8U8S32_IMPL_STR,
                _gemm_x8s8s32x_convolution_fwd_t<src_type, dst_type>);

        virtual status_t init() override {
            using namespace data_type;
            using namespace memory_format;

            assert(this->engine()->kind() == engine_kind::cpu);

            bool ok = true
                && this->set_default_params() == status::success
                && utils::one_of(this->desc()->prop_kind,
                        prop_kind::forward_training,
                        prop_kind::forward_inference)
                && utils::one_of(this->desc()->alg_kind,
                        alg_kind::convolution_auto,
                        alg_kind::convolution_direct)
                && !this->has_zero_dim_memory()
                && this->desc()->src_desc.data_type == src_type
                && this->desc()->dst_desc.data_type == dst_type
                && this->desc()->weights_desc.data_type == s8
                && IMPLICATION(this->with_bias(), utils::one_of(
                            this->desc()->bias_desc.data_type, f32, s32, s8,
                            u8))
                && this->desc()->accum_data_type == data_type::s32
                && utils::one_of(this->src_pd_.desc()->format, nhwc, ndhwc)
                && this->src_pd_.desc()->format == this->dst_pd_.desc()->format
                && IMPLICATION(this->src_pd_.desc()->format == nhwc,
                        this->weights_pd_.desc()->format == (this->with_groups()
                                ? ((src_type == data_type::s8) ? hwigo_s8s8 : hwigo)
                                : ((src_type == data_type::s8) ? hwio_s8s8 : hwio)))
                && IMPLICATION(this->src_pd_.desc()->format == ndhwc,
                        this->weights_pd_.desc()->format == (this->with_groups()
                                ? ((src_type == data_type::s8) ? dhwigo_s8s8 : dhwigo)
                                : ((src_type == data_type::s8) ? dhwio_s8s8 : dhwio)))
                && this->is_gemm_conv_format();
            if (!ok) return status::unimplemented;

            auto scratchpad = scratchpad_registry().registrar();
            return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
                    *this->desc(), this->src_pd(), this->weights_pd(0),
                    this->dst_pd(), mkldnn_get_max_threads());
        }

        jit_gemm_conv_conf_t jcp_;

    protected:
        memory_format_t src_format() const {
            using namespace memory_format;
            const size_t ndims_sp = this->desc()->src_desc.ndims - 4;
            return (utils::pick(ndims_sp, nhwc, ndhwc));
        }

        memory_format_t wei_format() const {
            using namespace memory_format;
            const size_t ndims_sp = this->desc()->src_desc.ndims - 4;
            return this->with_groups()
                ? (src_type == data_type::s8) ? utils::pick(ndims_sp, hwigo_s8s8, dhwigo_s8s8)
                                              : utils::pick(ndims_sp, hwigo, dhwigo)
                : (src_type == data_type::s8) ? utils::pick(ndims_sp, hwio_s8s8, dhwio_s8s8)
                                              : utils::pick(ndims_sp, hwio, dhwio);
        }

        virtual status_t set_default_params() override {
            using namespace memory_format;
            if (this->src_pd_.desc()->format == any)
                CHECK(this->src_pd_.set_format(src_format()));
            if (this->dst_pd_.desc()->format == any)
                CHECK(this->dst_pd_.set_format(src_format()));
            if (this->weights_pd_.desc()->format == any)
                CHECK(this->weights_pd_.set_format(wei_format()));
            if (this->bias_pd_.desc()->format == any)
                CHECK(this->bias_pd_.set_format(x));
            if (this->desc()->alg_kind == alg_kind::convolution_auto)
                CHECK(this->set_alg_kind(alg_kind::convolution_direct));
            return status::success;
        }

        virtual bool is_gemm_conv_format() const {
            using namespace mkldnn::impl::primitive_kind;
            auto const &po = this->attr()->post_ops_;
            auto is_relu = [&](int idx) {
                return po.entry_[idx].is_relu(true, false); };

            switch (po.len_) {
            case 0: return true;
            case 1: return is_relu(0) || po.contain(sum, 0);
            case 2: return po.contain(sum, 0) && is_relu(1);
            default: return false;
            }
        }
    };

    _gemm_x8s8s32x_convolution_fwd_t(const pd_t *apd, const input_vector &inputs,
           const output_vector &outputs)
        : cpu_primitive_t(apd, inputs, outputs, true) {
        pp_ker_ = new pp_ker_t(this->pd());
    }
    ~_gemm_x8s8s32x_convolution_fwd_t() {
        delete pp_ker_;
    }

    typedef typename prec_traits<src_type>::type src_data_t;
    typedef typename prec_traits<data_type::s8>::type wei_data_t;
    typedef typename prec_traits<dst_type>::type dst_data_t;
    typedef typename prec_traits<data_type::s32>::type acc_data_t;

    virtual void execute(event_t *e) const {
        execute_forward();
        e->set_state(event_t::ready);
    }

private:
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
    void execute_forward() const;
    // XXX: this is throwaway code that will become unnecessary when we have a
    // sufficiently advanced igemm jit generator that supports quantization,
    // relu, and whatnot
    class pp_ker_t : jit_generator {
    public:
        DECLARE_CPU_JIT_AUX_FUNCTIONS(
        _gemm_x8s8s32x_convolution_fwd_t::pp_kernel);
        pp_ker_t(const pd_t *pd);

        void operator()(dst_data_t *dst, const acc_data_t *acc,
            const char *bias, const float *scales,
            float nslope, float sum_scale, float signed_scale,
            int g, size_t start, size_t end);

        size_t dst_os_stride_;

    private:
        void generate();

        struct ker_args {
            dst_data_t *dst;
            const acc_data_t *acc;
            const char *bias;
            const float *scales;
            float nslope;
            float sum_scale;
            float signed_scale;
            size_t len;
            size_t oc_offset;
        };
        void(*ker_)(const ker_args *args);

        const jit_gemm_conv_conf_t jcp_;
        size_t OC_;
        size_t OS_;
        data_type_t bias_data_type_;
        size_t bias_data_type_size_;
        size_t scale_idx_mult_;
        round_mode_t rmode_;
        bool do_bias_;
        bool do_relu_;
        bool do_sum_;
        bool do_signed_scaling_;
        size_t vlen_;
    };


    void execute_forward_thr(const int ithr, const int nthr,
            const src_data_t *src_base, const wei_data_t *wei_base,
            const char *bia_base, dst_data_t *dst_base,
            const memory_tracking::grantor_t &scratchpad) const;

    int nthr_;
    pp_ker_t *pp_ker_;

};

template <data_type_t dst_type>
struct _gemm_u8s8s32x_convolution_bwd_data_t: public cpu_primitive_t {
    struct pd_t: public cpu_convolution_bwd_data_pd_t{
        pd_t(engine_t *engine,
                const convolution_desc_t *adesc, const primitive_attr_t *attr,
                const convolution_fwd_pd_t *hint_fwd_pd)
            : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd)
            , jcp_() {}

        DECLARE_COMMON_PD_T(IGEMM_S8U8S32_IMPL_STR,
                _gemm_u8s8s32x_convolution_bwd_data_t<dst_type>);

        virtual status_t init() override {
            using namespace data_type;
            using namespace memory_format;

            assert(this->engine()->kind() == engine_kind::cpu);

            bool ok = true
                && this->set_default_params() == status::success
                && this->desc()->prop_kind == prop_kind::backward_data
                && utils::one_of(this->desc()->alg_kind, alg_kind::convolution_auto,
                           alg_kind::convolution_direct)
                && !this->has_zero_dim_memory()
                && this->desc()->diff_src_desc.data_type == dst_type
                && this->desc()->diff_dst_desc.data_type == u8
                && this->desc()->weights_desc.data_type == s8
                && IMPLICATION(this->with_bias(), utils::one_of(
                            this->desc()->bias_desc.data_type, f32, s32, s8,
                            u8))
                && this->desc()->accum_data_type == data_type::s32
                && utils::everyone_is(nhwc, this->diff_src_pd_.desc()->format,
                        this->diff_dst_pd_.desc()->format)
                && this->weights_pd_.desc()->format == (this->with_groups()
                        ? hwigo : hwio)
                && attr()->post_ops_.has_default_values();
            if (!ok) return status::unimplemented;

            auto scratchpad = scratchpad_registry().registrar();
            return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
                    *this->desc(), this->diff_src_pd(), this->weights_pd(0),
                    this->diff_dst_pd(), mkldnn_get_max_threads());
        }

        virtual bool support_bias() const override { return true; }

        jit_gemm_conv_conf_t jcp_;

    protected:
        virtual status_t set_default_params() override {
            using namespace memory_format;

            if (this->diff_src_pd_.desc()->format == any)
                CHECK(this->diff_src_pd_.set_format(nhwc));
            if (this->diff_dst_pd_.desc()->format == any)
                CHECK(this->diff_dst_pd_.set_format(nhwc));
            if (this->weights_pd_.desc()->format == any)
                CHECK(this->weights_pd_.set_format(
                            this->with_groups() ? hwigo : hwio));
            if (bias_pd_.desc()->format == any)
                CHECK(bias_pd_.set_format(x));
            if (this->desc()->alg_kind == alg_kind::convolution_auto)
                CHECK(this->set_alg_kind(alg_kind::convolution_direct));
             return status::success;
        }
    };

    _gemm_u8s8s32x_convolution_bwd_data_t(const pd_t *apd, const input_vector &inputs,
           const output_vector &outputs)
        : cpu_primitive_t(apd, inputs, outputs, true) {}
    ~_gemm_u8s8s32x_convolution_bwd_data_t() {}

    typedef typename prec_traits<data_type::u8>::type diff_dst_data_t;
    typedef typename prec_traits<data_type::s8>::type wei_data_t;
    typedef typename prec_traits<dst_type>::type diff_src_data_t;
    typedef typename prec_traits<data_type::s32>::type acc_data_t;

    virtual void execute(event_t *e) const {
        execute_backward_data();
        e->set_state(event_t::ready);
    }

private:
    void execute_backward_data() const;
    void execute_backward_data_thr(const int ithr, const int nthr,
            const diff_dst_data_t *diff_dst_base, const wei_data_t *wei_base,
            const char *bia_base, diff_src_data_t *diff_src_base,
            const memory_tracking::grantor_t &scratchpad) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
};

}
}
}

#endif