onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation 
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
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346

/*******************************************************************************
* Copyright 2016 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 CPU_X64_JIT_AVX2_CONV_KERNEL_F32_HPP
#define CPU_X64_JIT_AVX2_CONV_KERNEL_F32_HPP

#include "common/c_types_map.hpp"
#include "common/memory.hpp"
#include "common/memory_tracking.hpp"

#include "cpu/x64/injectors/jit_uni_postops_injector.hpp"
#include "cpu/x64/jit_generator.hpp"
#include "cpu/x64/jit_primitive_conf.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace x64 {

struct jit_avx2_conv_fwd_kernel_f32_t : public jit_generator_t {
    jit_avx2_conv_fwd_kernel_f32_t(const jit_conv_conf_t &ajcp,
            const primitive_attr_t &attr, const memory_desc_t &dst_md);

    DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx2_conv_fwd_kernel_f32_t)

    static status_t init_conf(jit_conv_conf_t &jcp,
            const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
            const memory_desc_wrapper &weights_d,
            const memory_desc_wrapper &dst_d, const primitive_attr_t &attr);
    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
            const jit_conv_conf_t &jcp);

    jit_conv_conf_t jcp;
    const primitive_attr_t &attr_;

private:
    std::unique_ptr<injector::jit_uni_postops_injector_t<avx2>>
            postops_injector_;

    constexpr static int isa_simd_width_
            = cpu_isa_traits_t<avx2>::vlen / sizeof(float);
    using reg64_t = const Xbyak::Reg64;
    reg64_t reg_input = rax;
    reg64_t aux_reg_input = r8;
    reg64_t reg_kernel = rdx;
    reg64_t aux_reg_kernel = r9;
    reg64_t reg_output = rsi;
    reg64_t reg_bias = rbx;

    reg64_t aux_reg_inp_d = r11;
    reg64_t aux_reg_ker_d = abi_not_param1;

    reg64_t reg_ki = rsi;
    reg64_t kj = r10;
    reg64_t oi_iter = r11;
    reg64_t ki_iter = r12;
    reg64_t reg_channel = ki_iter;
    reg64_t reg_kh = abi_not_param1;
    reg64_t reg_oc_blocks = r14;
    reg64_t imm_addr64 = r15;
    reg64_t reg_long_offt = r15;
    Xbyak::Reg32 reg_ci_flag = r13d;
    Xbyak::Reg32 reg_oc_flag = r14d;

    /* binary post-ops operand */
    reg64_t temp_offset_reg = r12;

    Xbyak::Ymm ytmp = Xbyak::Ymm(14);

    inline void oh_step_unroll_kw(
            int ur_w, int pad_l, int pad_r, int oc_blocks);
    inline void oh_step_nopad(int ur_w, int pad_l, int pad_r, int oc_blocks);
    void apply_postops(const int oc_blocks, const int ur_w, const int oc_tail);
    inline void width_blk_step(int ur_w, int pad_l, int pad_r, int oc_blocks);
    inline void solve_common(int oc_blocks);

    inline dim_t filter_w_to_input(int ki, int oi = 0, int pad_l = 0) const {
        return static_cast<dim_t>(ki) * (jcp.dilate_w + 1) + oi * jcp.stride_w
                - pad_l;
    }
    inline dim_t filter_h_to_input(int ki) const {
        return static_cast<dim_t>(ki) * (jcp.dilate_h + 1) * jcp.iw;
    }
    inline dim_t filter_d_to_input(int ki) const {
        return static_cast<dim_t>(ki) * (jcp.dilate_d + 1) * jcp.iw * jcp.ih;
    }

    inline dim_t get_input_offset(int i_ic, int i_iw) const {
        dim_t offset;
        if (utils::one_of(jcp.src_tag, format_tag::ncw, format_tag::nchw,
                    format_tag::ncdhw)) {
            offset = static_cast<dim_t>(i_ic) * jcp.id * jcp.ih * jcp.iw + i_iw;
        } else if (utils::one_of(jcp.src_tag, format_tag::nwc, format_tag::nhwc,
                           format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_iw) * jcp.ic * jcp.ngroups + i_ic;
        } else {
            offset = static_cast<dim_t>(i_iw) * jcp.ic_block + i_ic;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_output_offset(int i_oc_block, int i_ow) const {
        dim_t offset;
        if (utils::one_of(jcp.dst_tag, format_tag::nwc, format_tag::nhwc,
                    format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_ow) * jcp.oc * jcp.ngroups
                    + i_oc_block * jcp.oc_block;
        } else {
            offset = static_cast<dim_t>(i_oc_block) * jcp.od * jcp.oh * jcp.ow
                            * jcp.oc_block
                    + i_ow * jcp.oc_block;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_kernel_offset(int i_oc_block, int ki, int i_ic) const {
        dim_t block_step_size = static_cast<dim_t>(jcp.ic_block) * jcp.oc_block;
        dim_t ic_block_step_size = static_cast<dim_t>(jcp.kd) * jcp.kh * jcp.kw
                * block_step_size;
        dim_t oc_block_step_size
                = static_cast<dim_t>(jcp.nb_ic) * ic_block_step_size;
        dim_t offset = static_cast<dim_t>(i_oc_block) * oc_block_step_size
                + static_cast<dim_t>(ki) * block_step_size
                + static_cast<dim_t>(i_ic) * jcp.oc_block;
        return sizeof(float) * offset;
    }

    inline bool is_src_layout_nxc() const {
        return utils::one_of(jcp.src_tag, format_tag::ndhwc, format_tag::nhwc,
                format_tag::nwc);
    }

    void generate() override;
};

struct jit_avx2_conv_bwd_data_kernel_f32_t : public jit_generator_t {
    DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx2_conv_bwd_data_kernel_f32_t)

    jit_avx2_conv_bwd_data_kernel_f32_t(const jit_conv_conf_t &ajcp)
        : jit_generator_t(jit_name()), jcp(ajcp) {}

    static status_t init_conf(jit_conv_conf_t &jcp,
            const convolution_desc_t &cd, const memory_desc_wrapper &diff_src_d,
            const memory_desc_wrapper &weights_d,
            const memory_desc_wrapper &diff_dst_d);
    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
            const jit_conv_conf_t &jcp);

    jit_conv_conf_t jcp;

private:
    using reg64_t = const Xbyak::Reg64;

    reg64_t reg_ddst = rax;
    reg64_t aux_reg_ddst = r8;
    reg64_t reg_kernel = rdx;
    reg64_t aux_reg_kernel = r10;
    reg64_t reg_dsrc = rsi;
    reg64_t aux_reg_ddst_oc_loop = rbx; // used in ndims < 5 case only
    reg64_t aux_reg_kernel_oc_loop = abi_not_param1; /* used in ndims < 5
                                                        case only */

    reg64_t aux_reg_dst_d = r12; // used in ndims == 5 case only
    reg64_t aux_reg_ker_d = r14; // used in ndims == 5 case only

    reg64_t reg_ki = abi_not_param1; // used in ndims == 5 case only
    reg64_t kj = r11;
    reg64_t oi_iter = r12;
    reg64_t reg_kh = r14;
    reg64_t reg_channel = r13; // used in ndims < 5 case only
    reg64_t reg_channel_work = r9; // used in ndims < 5 case only
    reg64_t reg_long_offt = r15;
    reg64_t reg_reduce_work = reg_long_offt;
    Xbyak::Reg32 reg_ci_flag = r13d; // used for nxc tails

    inline void compute_loop(int ur_w, int l_overflow, int r_overflow);

    void generate() override;

    inline int get_iw_start(int ki, int l_overflow) const {
        int res = (jcp.iw - 1 + jcp.r_pad) % jcp.stride_w
                + l_overflow * jcp.stride_w
                - (jcp.kw - 1 - ki) * (jcp.dilate_w + 1);
        while (res < 0)
            res += jcp.stride_w;

        return res;
    }

    inline int get_iw_end(int ur_w, int ki, int r_overflow) const {
        if (utils::one_of(ur_w, jcp.iw, jcp.ur_w_tail))
            ur_w += nstl::min(0, jcp.r_pad); // remove negative padding
        int res = (ur_w - 1 + jcp.l_pad) % jcp.stride_w
                + r_overflow * jcp.stride_w - ki * (jcp.dilate_w + 1);
        while (res < 0)
            res += jcp.stride_w;

        return ur_w - res;
    }

    inline dim_t filter_w_to_ddst(int ki, int oi = 0, int pad_l = 0) const {
        return (oi + pad_l - ki * (jcp.dilate_w + 1)) / jcp.stride_w;
    }

    inline dim_t get_ddst_offset(int i_oc_block, int i_ow, int i_oc) const {
        dim_t offset;
        if (utils::one_of(jcp.dst_tag, format_tag::nwc, format_tag::nhwc,
                    format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_ow) * jcp.oc * jcp.ngroups
                    + i_oc_block * jcp.oc_block + i_oc;
        } else {
            offset = static_cast<dim_t>(i_oc_block) * jcp.od * jcp.oh * jcp.ow
                            * jcp.oc_block
                    + i_ow * jcp.oc_block + i_oc;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_dsrc_offset(int i_ic_block, int i_iw) const {
        dim_t offset;
        if (utils::one_of(jcp.src_tag, format_tag::nwc, format_tag::nhwc,
                    format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_iw) * jcp.ic * jcp.ngroups
                    + i_ic_block * jcp.ic_block;
        } else {
            offset = static_cast<dim_t>(i_ic_block) * jcp.id * jcp.ih * jcp.iw
                            * jcp.ic_block
                    + i_iw * jcp.ic_block;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_kernel_offset(
            int i_oc_block, int i_ic_block, int ki, int i_oc) const {
        dim_t block_step_size = static_cast<dim_t>(jcp.ic_block) * jcp.oc_block;
        dim_t ic_block_step_size = static_cast<dim_t>(jcp.kd) * jcp.kh * jcp.kw
                * block_step_size;
        dim_t oc_block_step_size
                = static_cast<dim_t>(jcp.nb_ic) * ic_block_step_size;
        dim_t offset = static_cast<dim_t>(i_oc_block) * oc_block_step_size
                + i_ic_block * ic_block_step_size + ki * block_step_size
                + i_oc * jcp.ic_block;
        return sizeof(float) * offset;
    }
};

struct jit_avx2_conv_bwd_weights_kernel_f32_t : public jit_generator_t {
    DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx2_conv_bwd_weights_kernel_f32_t)

    jit_avx2_conv_bwd_weights_kernel_f32_t(const jit_conv_conf_t &ajcp)
        : jit_generator_t(jit_name()), jcp(ajcp) {}

    static status_t init_conf(jit_conv_conf_t &jcp,
            const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
            const memory_desc_wrapper &diff_weights_d,
            const memory_desc_wrapper &diff_dst_d);
    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
            const jit_conv_conf_t &jcp);

    jit_conv_conf_t jcp;

private:
    using reg64_t = const Xbyak::Reg64;
    reg64_t reg_input = rax;
    reg64_t reg_kernel = rdx;
    reg64_t reg_output = rsi;
    reg64_t b_ic = abi_not_param1;
    reg64_t kj = r8;
    reg64_t reg_kh = r9;
    reg64_t reg_ur_w_trips = r10;
    reg64_t reg_tmp = r11;
    reg64_t reg_oj = r15;
    reg64_t reg_ih_count = rbx;
    reg64_t aux_reg_input = r12;
    reg64_t aux_reg_kernel = r13;
    reg64_t ki = r14;
    reg64_t reg_long_offt = r11;
    reg64_t reg_channel = reg_ih_count; // used for nxc tails
    Xbyak::Reg32 reg_ci_flag = r9d; // used for nxc tails

    inline void od_step_comeback_pointers();
    inline void oh_step_comeback_pointers();
    inline void compute_ic_block_step(int ur_w, int pad_l, int pad_r,
            int ic_block_step, int input_offset, int kernel_offset,
            int output_offset);
    inline void compute_oh_step_disp();
    inline void compute_oh_step_unroll_ow(int ic_block_step, int max_ur_w);
    inline void compute_oh_step_common(int ic_block_step, int max_ur_w);
    inline void compute_oh_loop_common();

    inline dim_t get_input_offset(int i_ic, int i_iw) const {
        dim_t offset;
        if (utils::one_of(jcp.src_tag, format_tag::ncw, format_tag::nchw,
                    format_tag::ncdhw)) {
            offset = static_cast<dim_t>(i_ic) * jcp.id * jcp.ih * jcp.iw + i_iw;
        } else if (utils::one_of(jcp.src_tag, format_tag::nwc, format_tag::nhwc,
                           format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_iw) * jcp.ic * jcp.ngroups + i_ic;
        } else {
            offset = static_cast<dim_t>(i_iw) * jcp.ic_block + i_ic;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_output_offset(int i_oc_block, int i_ow) const {
        dim_t offset;
        if (utils::one_of(jcp.dst_tag, format_tag::nwc, format_tag::nhwc,
                    format_tag::ndhwc)) {
            offset = static_cast<dim_t>(i_ow) * jcp.oc * jcp.ngroups
                    + i_oc_block * jcp.oc_block;
        } else {
            offset = static_cast<dim_t>(i_oc_block) * jcp.od * jcp.oh * jcp.ow
                            * jcp.oc_block
                    + i_ow * jcp.oc_block;
        }
        return sizeof(float) * offset;
    }

    inline dim_t get_kernel_offset(int ki, int i_ic) const {
        dim_t block_step_size = static_cast<dim_t>(jcp.ic_block) * jcp.oc_block;
        dim_t offset = static_cast<dim_t>(ki) * block_step_size
                + i_ic * jcp.oc_block;
        return sizeof(float) * offset;
    }
    void generate() override;
};

} // namespace x64
} // namespace cpu
} // namespace impl
} // namespace dnnl

#endif