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
|
/*******************************************************************************
* Copyright 2017-2018 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 JIT_AVX512_COMMON_CONV_WINOGRAD_KERNEL_F32_HPP
#define JIT_AVX512_COMMON_CONV_WINOGRAD_KERNEL_F32_HPP
#include "c_types_map.hpp"
#include "cpu_memory.hpp"
#include "jit_generator.hpp"
#include "jit_primitive_conf.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
//alpha determines the output tile_size
constexpr int alpha = 6;
constexpr int tile_size = 4;
//simd length used for vectorization
constexpr int simd_w = 16;
struct _jit_avx512_common_conv_winograd_data_kernel_f32 : public jit_generator {
_jit_avx512_common_conv_winograd_data_kernel_f32(
jit_conv_winograd_conf_t ajcp)
: jcp(ajcp)
{
//******************* First iter kernel ********************//
this->gemm_loop_generate(true);
gemm_loop_ker_first_iter
= (decltype(gemm_loop_ker_first_iter)) this->getCode();
//************** Subsequent iterations kernel **************//
if (jcp.dimK_nb_block > 1) {
align();
const Xbyak::uint8 *addr = getCurr();
this->gemm_loop_generate(false);
gemm_loop_ker = (decltype(gemm_loop_ker))addr;
}
}
DECLARE_CPU_JIT_AUX_FUNCTIONS(_jit_avx512_common_conv_winograd_data_kernel_f32)
static status_t init_conf_common(jit_conv_winograd_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);
static status_t init_conf_kernel(
jit_conv_winograd_conf_t &jcp, int dimM, int dimN, int dimK);
jit_conv_winograd_conf_t jcp;
void (*gemm_loop_ker)(float *, const float *, const float *);
void (*gemm_loop_ker_first_iter)(float *, const float *, const float *);
protected:
using reg64_t = const Xbyak::Reg64;
enum { typesize = sizeof(float) };
void gemm_loop_generate(bool is_beta_zero);
/* registers used for GEMM */
reg64_t reg_dstC = abi_param1;
reg64_t reg_srcA = abi_param2;
reg64_t reg_srcB = abi_param3;
reg64_t reg_dimM_block_loop_cnt = r10;
reg64_t reg_dimK_block_loop_cnt = r11;
};
struct jit_avx512_common_conv_winograd_fwd_kernel_f32
: _jit_avx512_common_conv_winograd_data_kernel_f32 {
using _jit_avx512_common_conv_winograd_data_kernel_f32::
_jit_avx512_common_conv_winograd_data_kernel_f32;
static bool post_ops_ok(jit_conv_conf_t &jcp, const primitive_attr_t &attr);
static status_t init_conf(jit_conv_winograd_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,
bool with_relu = false, float relu_negative_slope = 0.);
};
struct jit_avx512_common_conv_winograd_bwd_data_kernel_f32
: public _jit_avx512_common_conv_winograd_data_kernel_f32 {
using _jit_avx512_common_conv_winograd_data_kernel_f32::
_jit_avx512_common_conv_winograd_data_kernel_f32;
static status_t init_conf(jit_conv_winograd_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);
};
struct jit_avx512_common_conv_winograd_bwd_weights_kernel_f32
: public jit_generator {
DECLARE_CPU_JIT_AUX_FUNCTIONS(_jit_avx512_common_conv_winograd_bwd_weights_kernel_f32)
jit_avx512_common_conv_winograd_bwd_weights_kernel_f32(
jit_conv_winograd_conf_t ajcp)
: jcp(ajcp)
{
//******************* First iter kernel ********************//
{
align();
const Xbyak::uint8 *addr = getCurr();
this->gemm_loop_generate(true);
gemm_loop_ker_first_iter = (decltype(gemm_loop_ker_first_iter))addr;
}
if (jcp.tile_block > 1) {
align();
const Xbyak::uint8 *addr = getCurr();
this->gemm_loop_generate(false);
gemm_loop_ker = (decltype(gemm_loop_ker))addr;
}
if (jcp.ver == ver_4fma) {
align();
const Xbyak::uint8 *addr = getCurr();
this->transpose_ker_generate();
transpose_4fma_ker = (decltype(transpose_4fma_ker))addr;
}
}
static status_t init_conf(jit_conv_winograd_conf_t &jcp,
const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
const memory_desc_wrapper &diff_dst_d,
const memory_desc_wrapper &diff_weights_d);
jit_conv_winograd_conf_t jcp;
void (*gemm_loop_ker)(float *, const float *, const float *);
void (*gemm_loop_ker_first_iter)(float *, const float *, const float *);
void (*transpose_4fma_ker)(float *, float *);
private:
using reg64_t = const Xbyak::Reg64;
enum { typesize = sizeof(float) };
void gemm_loop_generate(bool is_first_tile);
void transpose_ker_generate();
reg64_t reg_origB = abi_param2;
reg64_t reg_transB = abi_param1;
reg64_t reg_dstC = abi_param1;
reg64_t reg_srcA_const = abi_param2;
reg64_t reg_srcB = abi_param3;
reg64_t reg_sp = rsp;
reg64_t reg_srcA = r9;
reg64_t reg_nb_ic = r10;
reg64_t reg_loop_cpt = r11;
reg64_t reg_transB_idx = r13;
/* Registers used by new kernel */
reg64_t reg_dimM_block_loop_cnt = r10;
reg64_t reg_dimK_block_loop_cnt = r12;
reg64_t reg_dimN_block_loop_cnt = r11;
};
}
}
}
#endif
|
