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

/*******************************************************************************
* Copyright 2026 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 GPU_INTEL_GATED_MLP_REF_HPP
#define GPU_INTEL_GATED_MLP_REF_HPP

#include "common/c_types_map.hpp"
#include "common/gated_mlp_pd.hpp"
#include "common/matmul_pd.hpp"
#include "common/primitive.hpp"
#include "common/utils.hpp"

#include "gpu/gpu_resource.hpp"
#include "gpu/intel/compute/kernel.hpp"
#include "gpu/intel/primitive.hpp"
#include "gpu/intel/primitive_conf.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace gated_mlp {

struct ref_t : public primitive_t {
    using primitive_t::primitive_t;
    struct pd_t : public gated_mlp_pd_t {
        using gated_mlp_pd_t::gated_mlp_pd_t;

        DECLARE_COMMON_PD_T("ocl:ref:any", ref_t);

        status_t init(impl::engine_t *engine) {
            VDISPATCH_GATED_MLP(pd_ok(), VERBOSE_INCONSISTENT_PRB);
            VDISPATCH_GATED_MLP(set_default_formats(), VERBOSE_UNSUPPORTED_TAG);

            memory_desc_t gate_dst_md, up_dst_md;
            CHECK(get_gate_dst_md(gate_dst_md));
            CHECK(get_up_dst_md(up_dst_md));

            primitive_attr_t gate_attr;
            CHECK(move_attr(
                    gate_attr, DNNL_ARG_WEIGHTS_GATE, DNNL_ARG_WEIGHTS));
            CHECK(gate_attr.post_ops_.append_eltwise(
                    1.f, activation(), 1.f, 0.f));
            CHECK(gate_attr.post_ops_.append_binary(
                    alg_kind::binary_mul, &up_dst_md));
            VDISPATCH_GATED_MLP_SC(
                    create_matmul(gemm_gate_pd_, engine, gate_attr,
                            arg_md(DNNL_ARG_SRC), arg_md(DNNL_ARG_WEIGHTS_GATE),
                            &gate_dst_md),
                    "internal error in gemm_gate_pd");

            primitive_attr_t up_attr;
            CHECK(move_attr(up_attr, DNNL_ARG_WEIGHTS_UP, DNNL_ARG_WEIGHTS));
            VDISPATCH_GATED_MLP_SC(
                    create_matmul(gemm_up_pd_, engine, up_attr,
                            arg_md(DNNL_ARG_SRC), arg_md(DNNL_ARG_WEIGHTS_UP),
                            &up_dst_md),
                    "internal error in gemm_up_pd");

            primitive_attr_t down_attr;
            CHECK(move_attr(
                    down_attr, DNNL_ARG_WEIGHTS_DOWN, DNNL_ARG_WEIGHTS));
            VDISPATCH_GATED_MLP_SC(
                    create_matmul(gemm_down_pd_, engine, down_attr,
                            &gate_dst_md, arg_md(DNNL_ARG_WEIGHTS_DOWN),
                            arg_md(DNNL_ARG_DST)),
                    "internal error in gemm_down_pd");

            using namespace memory_tracking::names;
            auto scratchpad = scratchpad_registry().registrar();
            scratchpad.book(key_matmul_wei_trans,
                    memory_desc_wrapper(gate_dst_md).size(), 1,
                    OCL_BUFFER_ALIGNMENT);
            scratchpad.book(key_matmul_src_trans,
                    memory_desc_wrapper(up_dst_md).size(), 1,
                    OCL_BUFFER_ALIGNMENT);
            scratchpad.book(key_nested_multiple + DNNL_ARG_WEIGHTS_GATE,
                    gemm_gate_pd_->scratchpad_registry());
            scratchpad.book(key_nested_multiple + DNNL_ARG_WEIGHTS_UP,
                    gemm_up_pd_->scratchpad_registry());
            scratchpad.book(key_nested_multiple + DNNL_ARG_WEIGHTS_DOWN,
                    gemm_down_pd_->scratchpad_registry());
            return status::success;
        }

        status_t get_gate_dst_md(memory_desc_t &retn) const {
            data_type_t dt = arg_md(DNNL_ARG_WEIGHTS_DOWN)->data_type;
            switch (dt) {
                case data_type::f32:
                case data_type::bf16:
                case data_type::f16: break;
                case data_type::u4:
                case data_type::s4:
                case data_type::u8:
                case data_type::s8:
                    dt = arg_md(DNNL_ARG_SRC)->data_type;
                    if (!utils::one_of(dt, data_type::bf16, data_type::f16))
                        dt = data_type::f16;
                    break;
                default: return status::unimplemented;
            }
            return get_gate_up_dst_md(retn, dt);
        }

        status_t get_up_dst_md(memory_desc_t &retn) const {
            return get_gate_up_dst_md(retn, data_type::f32);
        }

        std::shared_ptr<primitive_desc_t> gemm_gate_pd_, gemm_up_pd_,
                gemm_down_pd_;

    private:
        status_t get_gate_up_dst_md(memory_desc_t &retn, data_type_t dt) const {
            std::vector<dim_t> dims {MB(), OC()};
            CHECK(memory_desc_init_by_tag(
                    retn, int(dims.size()), dims.data(), dt, format_tag::ab));
            return status::success;
        }

        status_t move_attr(primitive_attr_t &retn, int from, int to) {
            CHECK(retn.set_fpmath_mode(
                    attr()->fpmath_.mode_, attr()->fpmath_.apply_to_int_));
            CHECK(retn.scales_.set(to, attr()->scales_.get(from)));
            CHECK(retn.zero_points_.set(to, attr()->zero_points_.get(from)));
            return status::success;
        }

        status_t create_matmul(std::shared_ptr<primitive_desc_t> &retn,
                impl::engine_t *e, const primitive_attr_t &attr,
                const memory_desc_t *src_desc, const memory_desc_t *wei_desc,
                const memory_desc_t *dst_desc) {
            auto desc = matmul_desc_t();
            CHECK(impl::matmul_desc_init(
                    &desc, src_desc, wei_desc, nullptr, dst_desc));
            primitive_desc_iterator_t it(e, (op_desc_t *)&desc, &attr, nullptr);
            if (!it.is_initialized()) return status::out_of_memory;
            retn = *(++it);
            return (retn) ? status::success : status::unimplemented;
        }
    };

    status_t init(impl::engine_t *engine) override {
        CHECK(create_nested_primitive(gemm_gate_, pd()->gemm_gate_pd_, engine));
        CHECK(create_nested_primitive(gemm_up_, pd()->gemm_up_pd_, engine));
        CHECK(create_nested_primitive(gemm_down_, pd()->gemm_down_pd_, engine));
        return status::success;
    }

    status_t execute(const exec_ctx_t &ctx) const override {
        auto prep_weights_and_run
                = [&](exec_args_t &&args, int idx,
                          const std::shared_ptr<impl::primitive_t> &prim) {
            args[DNNL_ARG_WEIGHTS] = ctx.args().at(idx);
            if (!pd()->attr()->scales_.has_default_values(idx))
                args[DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS]
                        = ctx.args().at(DNNL_ARG_ATTR_SCALES | idx);
            if (!pd()->attr()->zero_points_.has_default_values(idx))
                args[DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_WEIGHTS]
                        = ctx.args().at(DNNL_ARG_ATTR_ZERO_POINTS | idx);
            exec_ctx_t nested_ctx(ctx, std::move(args));
            auto *nested_grantor
                    = create_nested_grantor(ctx.get_scratchpad_grantor(),
                            memory_tracking::names::key_nested_multiple + idx,
                            prim->pd()->scratchpad_registry());
            nested_ctx.set_scratchpad_grantor(nested_grantor);
            CHECK(prim->execute(nested_ctx));
            return status::success;
        };
        memory_desc_t gate_dst_md, up_dst_md;
        CHECK(pd()->get_gate_dst_md(gate_dst_md));
        CHECK(pd()->get_up_dst_md(up_dst_md));

        std::unique_ptr<memory_t, memory_deleter_t> inter_src_mem;
        auto inter_src_stor = ctx.get_scratchpad_grantor().get_memory_storage(
                memory_tracking::names::key_matmul_src_trans);
        CHECK(safe_ptr_assign(inter_src_mem,
                new memory_t(ctx.stream()->engine(), &up_dst_md,
                        std::move(inter_src_stor))));

        std::unique_ptr<memory_t, memory_deleter_t> inter_wei_mem;
        auto inter_wei_stor = ctx.get_scratchpad_grantor().get_memory_storage(
                memory_tracking::names::key_matmul_wei_trans);
        CHECK(safe_ptr_assign(inter_wei_mem,
                new memory_t(ctx.stream()->engine(), &gate_dst_md,
                        std::move(inter_wei_stor))));

        do {
            exec_args_t args;
            args[DNNL_ARG_SRC] = ctx.args().at(DNNL_ARG_SRC);
            args[DNNL_ARG_DST] = memory_arg_t {inter_src_mem.get(), false};
            CHECK(prep_weights_and_run(
                    std::move(args), DNNL_ARG_WEIGHTS_UP, gemm_up_));
        } while (false);
        do {
            exec_args_t args;
            args[DNNL_ARG_SRC] = ctx.args().at(DNNL_ARG_SRC);
            args[DNNL_ARG_DST] = memory_arg_t {inter_wei_mem.get(), false};
            // N.B.: not POST_OP(0), since POST_OP(0) is occupied by activation
            args[DNNL_ARG_ATTR_MULTIPLE_POST_OP(1) | DNNL_ARG_SRC_1]
                    = memory_arg_t {inter_src_mem.get(), true};
            CHECK(prep_weights_and_run(
                    std::move(args), DNNL_ARG_WEIGHTS_GATE, gemm_gate_));
        } while (false);
        do {
            exec_args_t args;
            args[DNNL_ARG_SRC] = memory_arg_t {inter_wei_mem.get(), true};
            args[DNNL_ARG_DST] = ctx.args().at(DNNL_ARG_DST);
            CHECK(prep_weights_and_run(
                    std::move(args), DNNL_ARG_WEIGHTS_DOWN, gemm_down_));
        } while (false);

        return status::success;
    }

private:
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
    std::shared_ptr<impl::primitive_t> gemm_gate_, gemm_up_, gemm_down_;
};

} // namespace gated_mlp
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif