1/*******************************************************************************
2* Copyright 2019-2022 Intel Corporation
3*
4* Licensed under the Apache License, Version 2.0 (the "License");
5* you may not use this file except in compliance with the License.
6* You may obtain a copy of the License at
7*
8* http://www.apache.org/licenses/LICENSE-2.0
9*
10* Unless required by applicable law or agreed to in writing, software
11* distributed under the License is distributed on an "AS IS" BASIS,
12* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13* See the License for the specific language governing permissions and
14* limitations under the License.
15*******************************************************************************/
16
17#ifndef CPU_X64_JIT_UNI_SOFTMAX_HPP
18#define CPU_X64_JIT_UNI_SOFTMAX_HPP
19
20#include <assert.h>
21
22#include "common/c_types_map.hpp"
23#include "common/memory_tracking.hpp"
24#include "common/primitive.hpp"
25#include "common/type_helpers.hpp"
26#include "common/utils.hpp"
27
28#include "cpu/cpu_softmax_pd.hpp"
29#include "cpu/x64/cpu_isa_traits.hpp"
30#include "cpu/x64/jit_avx512_core_bf16cvt.hpp"
31
32namespace dnnl {
33namespace impl {
34namespace cpu {
35namespace x64 {
36
37namespace softmax_impl {
38template <cpu_isa_t isa>
39struct driver_t;
40}
41
42template <cpu_isa_t isa>
43struct jit_uni_softmax_fwd_t : public primitive_t {
44 struct pd_t : public cpu_softmax_fwd_pd_t {
45 using cpu_softmax_fwd_pd_t::cpu_softmax_fwd_pd_t;
46
47 DECLARE_COMMON_PD_T(
48 JIT_IMPL_NAME_HELPER("jit:", isa, ""), jit_uni_softmax_fwd_t);
49
50 status_t init(engine_t *engine) {
51 auto is_dense = [&]() {
52 const memory_desc_wrapper src_d(src_md());
53 const auto &bd = src_d.blocking_desc();
54
55 if (!src_d.is_dense(true) || !src_d.only_padded_dim(axis()))
56 return false;
57
58 if (src_d.is_plain()) return bd.strides[axis()] == 1;
59
60 // It is fine to use float here as the kernel uses halfs of
61 // vector registers.
62 const auto blk_size = cpu_isa_traits<isa>::vlen / sizeof(float);
63 // 31 is a general limit, 2 is for unroll_regs_ = 4;
64 const size_t max_stride = (1LL << (31 - 2)) - 1;
65 const int last_blk = bd.inner_nblks - 1;
66 return bd.inner_blks[last_blk] == blk_size
67 && bd.inner_idxs[last_blk] == axis()
68 && sizeof(float) * bd.strides[axis()] < max_stride;
69 };
70
71 using namespace data_type;
72 using skip_mask_t = primitive_attr_t::skip_mask_t;
73
74 const auto src_dt = src_md()->data_type;
75 const auto dst_dt = dst_md()->data_type;
76 bool ok = mayiuse(isa) && is_fwd() && !has_zero_dim_memory()
77 && utils::one_of(src_dt, f32, bf16, f16, s8, u8)
78 && utils::one_of(dst_dt, f32, bf16, f16, s8, u8)
79 // s8/u8 are temporary limitations due to priorities
80 && IMPLICATION(
81 (utils::one_of(s8, src_dt, dst_dt)
82 || utils::one_of(u8, src_dt, dst_dt)),
83 is_superset(isa, avx512_core))
84 && IMPLICATION(utils::one_of(bf16, src_dt, dst_dt),
85 (is_superset(isa, avx512_core)
86 || (isa == avx2 && mayiuse(avx2_vnni_2))))
87 // for f16 we reuse avx512_core/avx2 just to avoid
88 // additional instantiation. Possible because we do not
89 // currently support post-ops for this primitive
90 && IMPLICATION(utils::one_of(f16, src_dt, dst_dt),
91 (is_superset(isa, avx512_core)
92 && mayiuse(avx512_core_fp16))
93 || (isa == avx2 && mayiuse(avx2_vnni_2)))
94 && attr()->has_default_values(skip_mask_t::scales_runtime)
95 && attr_scales_ok()
96 && set_default_formats() == status::success;
97 if (!ok) return status::unimplemented;
98
99 ok = memory_desc_wrapper(src_md()).similar_to(
100 memory_desc_wrapper(dst_md()), true, false, 0)
101 && is_dense(); // not dense impl can be easily done
102 if (!ok) return status::unimplemented;
103
104 // AVX2 only supports xf16 on plain layout now
105 ok = IMPLICATION(isa == avx2 && mayiuse(avx2_vnni_2)
106 && (utils::one_of(bf16, src_dt, dst_dt)
107 || utils::one_of(f16, src_dt, dst_dt)),
108 memory_desc_wrapper(src_md()).is_plain());
109 if (!ok) return status::unimplemented;
110
111 nthr_ = dnnl_get_max_threads();
112 init_scratchpad();
113
114 return status::success;
115 };
116
117 int nthr_; // To not exceed the limit in execute used for set up.
118
119 private:
120 void init_scratchpad() {
121 if (utils::one_of(
122 dst_md()->data_type, data_type::u8, data_type::s8)) {
123 auto scratchpad = scratchpad_registry().registrar();
124 scratchpad.template book<char>(
125 memory_tracking::names::key_softmax_interim_store,
126 axis_size(true) * sizeof(float) * nthr_);
127 }
128 }
129 };
130
131 jit_uni_softmax_fwd_t(const pd_t *apd);
132 ~jit_uni_softmax_fwd_t();
133
134 status_t init(engine_t *engine) override;
135
136 status_t execute(const exec_ctx_t &ctx) const override;
137
138private:
139 const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
140 softmax_impl::driver_t<isa> *softmax_driver_;
141};
142
143template <cpu_isa_t isa>
144struct jit_uni_softmax_bwd_t : public primitive_t {
145 struct pd_t : public cpu_softmax_bwd_pd_t {
146 using cpu_softmax_bwd_pd_t::cpu_softmax_bwd_pd_t;
147
148 DECLARE_COMMON_PD_T(
149 JIT_IMPL_NAME_HELPER("jit:", isa, ""), jit_uni_softmax_bwd_t);
150
151 status_t init(engine_t *engine) {
152 auto is_dense = [&]() {
153 const memory_desc_wrapper dst_d(dst_md());
154 const auto &bd = dst_d.blocking_desc();
155
156 if (!dst_d.is_dense(true) || !dst_d.only_padded_dim(axis()))
157 return false;
158
159 // It is fine to use float here as the kernel uses halfs of
160 // vector registers.
161 const auto blk_size = cpu_isa_traits<isa>::vlen / sizeof(float);
162 if (dst_d.is_plain())
163 return bd.strides[axis()] == 1;
164 else {
165 // 31 is a general limit, 2 is for unroll_regs_ = 4;
166 const size_t max_stride = (1LL << (31 - 2)) - 1;
167 const int last_blk = bd.inner_nblks - 1;
168 return bd.inner_blks[last_blk] == blk_size
169 && bd.inner_idxs[last_blk] == axis()
170 && sizeof(float) * bd.strides[axis()] < max_stride;
171 }
172 };
173
174 using namespace data_type;
175 bool ok = mayiuse(isa) && !is_fwd() && !has_zero_dim_memory()
176 && utils::one_of(dst_md()->data_type, f32, bf16, f16)
177 && utils::one_of(diff_dst_md()->data_type, f32, bf16, f16)
178 && utils::one_of(diff_src_md()->data_type, f32, bf16, f16)
179 && IMPLICATION(utils::one_of(bf16, dst_md()->data_type,
180 diff_dst_md()->data_type,
181 diff_src_md()->data_type),
182 is_superset(isa, avx512_core))
183 // for f16 we reuse avx512_core just to avoid additional
184 // instantiation.
185 && IMPLICATION(utils::one_of(f16, dst_md()->data_type,
186 diff_dst_md()->data_type,
187 diff_src_md()->data_type),
188 is_superset(isa, avx512_core)
189 && mayiuse(avx512_core_fp16))
190 && attr()->has_default_values()
191 && set_default_formats() == status::success;
192 if (!ok) return status::unimplemented;
193
194 ok = memory_desc_wrapper(diff_src_md())
195 .similar_to(memory_desc_wrapper(diff_dst_md()),
196 true, false, 0)
197 && memory_desc_wrapper(diff_dst_md())
198 == memory_desc_wrapper(dst_md())
199 && is_dense(); // not dense impl can be easily done
200 if (!ok) return status::unimplemented;
201
202 return status::success;
203 };
204 };
205
206 jit_uni_softmax_bwd_t(const pd_t *apd);
207 ~jit_uni_softmax_bwd_t();
208
209 status_t init(engine_t *engine) override;
210
211 status_t execute(const exec_ctx_t &ctx) const override;
212
213private:
214 const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
215
216 softmax_impl::driver_t<isa> *softmax_driver_;
217};
218
219} // namespace x64
220} // namespace cpu
221} // namespace impl
222} // namespace dnnl
223
224#endif
225
226// vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s
227