1/*******************************************************************************
2* Copyright 2021-2022 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_MATMUL_REF_MATMUL_INT8_HPP
18#define CPU_MATMUL_REF_MATMUL_INT8_HPP
19
20#include <assert.h>
21
22#include "common/c_types_map.hpp"
23#include "common/primitive.hpp"
24#include "common/type_helpers.hpp"
25#include "common/utils.hpp"
26
27#include "cpu/platform.hpp"
28#include "cpu/primitive_attr_postops.hpp"
29
30#include "cpu/matmul/cpu_matmul_pd.hpp"
31
32namespace dnnl {
33namespace impl {
34namespace cpu {
35namespace matmul {
36
37struct ref_matmul_int8_t : public primitive_t {
38 struct pd_t : public cpu_matmul_pd_t {
39 using cpu_matmul_pd_t::cpu_matmul_pd_t;
40
41 DECLARE_COMMON_PD_T("ref:any", ref_matmul_int8_t);
42
43 status_t init(engine_t *engine) {
44 using namespace data_type;
45 using smask_t = primitive_attr_t::skip_mask_t;
46 const auto src_type = src_md(0)->data_type;
47 const auto wei_type = weights_md(0)->data_type;
48 const auto bia_type = weights_md(1)->data_type;
49 const auto dst_type = dst_md(0)->data_type;
50
51 bool ok = utils::one_of(src_type, s8, u8) && wei_type == s8
52 && IMPLICATION(with_bias(),
53 utils::one_of(bia_type, f32, bf16, s32, s8, u8))
54 && utils::one_of(dst_type, f32, bf16, s32, s8, u8)
55 && attr()->has_default_values(smask_t::scales_runtime
56 | smask_t::zero_points_runtime
57 | smask_t::post_ops | smask_t::sum_dt,
58 dst_type)
59 && attr_.post_ops_.check_sum_consistent_dt(dst_type)
60 && attr_scales_ok() && attr_zero_points_ok()
61 && set_default_formats()
62 && attr_.set_default_formats(dst_md(0)) == status::success;
63 return ok ? status::success : status::unimplemented;
64 }
65
66 private:
67 bool attr_scales_ok() {
68 const std::vector<int> supported_args
69 = {DNNL_ARG_SRC, DNNL_ARG_WEIGHTS, DNNL_ARG_DST};
70 bool ok = attr()->scales_.has_default_values(supported_args);
71 for (int arg : supported_args) {
72 const auto &mask = attr()->scales_.get(arg).mask_;
73 if (arg == DNNL_ARG_WEIGHTS)
74 ok = ok
75 && (mask == 0
76 || mask == (1 << (dst_md()->ndims - 1)));
77 else
78 ok = ok && (mask == 0);
79 }
80 return ok;
81 }
82
83 bool attr_zero_points_ok() const {
84 int mask_src = 0, mask_wei = 0, mask_dst = 0;
85 attr()->zero_points_.get(DNNL_ARG_SRC, &mask_src);
86 attr()->zero_points_.get(DNNL_ARG_WEIGHTS, &mask_wei);
87 attr()->zero_points_.get(DNNL_ARG_DST, &mask_dst);
88
89 return (mask_src == 0 || mask_src == 1 << 1) && (mask_wei == 0)
90 && (mask_dst == 0 || mask_dst == 1 << 1);
91 }
92 };
93
94 ref_matmul_int8_t(const pd_t *apd) : primitive_t(apd) {}
95
96 status_t init(engine_t *engine) override {
97 ref_post_ops
98 = utils::make_unique<ref_post_ops_t>(pd()->attr()->post_ops_);
99 if (!ref_post_ops) return status::out_of_memory;
100 return status::success;
101 }
102
103 status_t execute(const exec_ctx_t &ctx) const override {
104 return execute_ref(ctx);
105 }
106
107private:
108 const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
109 status_t execute_ref(const exec_ctx_t &ctx) const;
110 std::unique_ptr<ref_post_ops_t> ref_post_ops;
111};
112
113} // namespace matmul
114} // namespace cpu
115} // namespace impl
116} // namespace dnnl
117
118#endif
119