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

/*******************************************************************************
* Copyright 2024 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.
*******************************************************************************/

#include "graph/backend/dnnl/kernels/sdp_primitive_config.hpp"
#include "graph/backend/dnnl/fusion_info.hpp"

#include "common/compiler_workarounds.hpp"

#define VCHECK_SDP_PRIMITIVE(cond, status, msg, ...) \
    VCONDCHECK(graph, create, check, sdp_primitive_kernel_t, (cond), status, \
            msg, ##__VA_ARGS__);

namespace dnnl {
namespace impl {
namespace graph {
namespace dnnl_impl {

op_ptr sdp_primitive_config_t::get_post_op(const op_ptr &op) const {
    const auto out_val = op->get_output_value(0);
    const auto &consumers = out_val->get_consumers();
    if (consumers.size() != 1) return nullptr;
    return consumers[0].get_op().shared_from_this();
}

status_t sdp_primitive_config_t::initial_check(
        const std::shared_ptr<subgraph_t> &sg,
        const std::vector<logical_tensor_t> &inputs,
        const std::vector<logical_tensor_t> &outputs) {
    // At least 3 inputs: Q, K, V
    VCHECK_SDP_PRIMITIVE(inputs.size() >= 3, status::invalid_arguments,
            "At least 3 inputs are required");

    const bool is_f32 = inputs[0].data_type == data_type::f32;
    bool has_genindex = false;

    // Dispatch f32 implicit causal mask cases into the f32 ukernel impl.
    for (auto &cur_op : sg->get_ops()) {
        const auto opk = cur_op->get_kind();
        // SDPA with static quantization and dequantization is currently
        // unsupported in the ukernel. Only compressed KV SDPA pattern with
        // dynamic dequantization before K and V is supported.
        VCHECK_SDP_PRIMITIVE(opk != graph::op_kind::Dequantize
                        && opk != graph::op_kind::Quantize,
                status::unimplemented, "Not support quantized SDPA");
        if (opk == graph::op_kind::GenIndex) { has_genindex = true; }
    }

    // step1(pattern check): Not support sdpa variants with select as mask
    // We already have a pattern matcher to ensure that the sdpa patterns
    // dispatch to here are knows ones, and we have quant check in sdpa base
    // kernel, so here we only check specific variants based on support matrix.
    const std::unordered_set<graph::op_kind_t> mm1_post_op_kind
            = {graph::op_kind::Divide, graph::op_kind::Multiply,
                    graph::op_kind::Add, graph::op_kind::Select,
                    graph::op_kind::SoftMax};
    op_ptr mm1 = nullptr, mm2 = nullptr, scale = nullptr;
    bool f32_inter = true;

    for (const auto &cur_op : sg->get_ops()) {
        const auto &op_kind = cur_op->get_kind();
        if (op_kind == graph::op_kind::DynamicDequantize
                && cur_op->get_attr<std::string>(op_attr::qtype)
                        == "per_group") {
            if (!cur_op->has_attr(op_attr::group_shape))
                return status::invalid_arguments;
            is_compressed_ = true;
            const auto &group_shape = cur_op->get_attr<std::vector<int64_t>>(
                    op_attr::group_shape);
            const auto &input_lt = cur_op->get_input_logical_tensor(0);
            if (static_cast<int>(group_shape.size()) != ltw(input_lt).ndims())
                return status::invalid_arguments;
            // TODO(zhitao): execute the reorder for scale and zps mannually if the
            // transpose attribute is specified as true.
            auto post_op = get_post_op(cur_op);
            if (post_op && post_op->get_kind() == graph::op_kind::MatMul
                    && post_op->has_attr(op_attr::transpose_b)
                    && post_op->get_attr<bool>(op_attr::transpose_b))
                return status::unimplemented;
        }
        if (op_kind != graph::op_kind::MatMul) continue;
        auto post_op = get_post_op(cur_op);
        if (post_op && mm1_post_op_kind.count(post_op->get_kind())) {
            mm1 = cur_op;
            const auto &lt_score = mm1->get_output_logical_tensor(0);
            f32_inter = f32_inter
                    && (ltw(lt_score).data_type() == data_type::f32);
            // Not support select between mm1 and scale(optional)
            // GPT-J:[mm1] --> [select] --> [scale]* --> [mask]* --> ...
            VCHECK_SDP_PRIMITIVE(post_op->get_kind() != graph::op_kind::Select,
                    status::unimplemented,
                    "Not support select between mm1 and scale(optional)");
            // scale
            if (post_op->get_kind() == graph::op_kind::Divide
                    || post_op->get_kind() == graph::op_kind::Multiply) {
                // Scale exists, update post_op and traverse to next op
                scale = post_op;
                post_op = get_post_op(post_op);
                const auto &lt_ss = scale->get_output_logical_tensor(0);
                f32_inter = f32_inter
                        && (ltw(lt_ss).data_type() == data_type::f32);
            }
            // mask
            if (post_op) {
                if (post_op->get_kind() == graph::op_kind::Add) {
                    // Mask exists, update post_op and traverse to next op
                    const auto mask = post_op;
                    const auto &lt_ms = mask->get_output_logical_tensor(0);
                    f32_inter = f32_inter
                            && (ltw(lt_ms).data_type() == data_type::f32);
                    post_op = get_post_op(post_op);
                }
                // Not support select after scale(optional) and mask(optional)
                // Distill-Bert:[mm1] --> [scale]* --> [mask]* --> [select] --> ...
                VCHECK_SDP_PRIMITIVE(post_op
                                && post_op->get_kind()
                                        != graph::op_kind::Select,
                        status::unimplemented,
                        "Not support select after scale(optional) and "
                        "mask(optional)");
            }

            if (post_op) {
                if (post_op->get_kind() == graph::op_kind::SoftMax) {
                    const auto &softmax = post_op;
                    softmax_mode_
                            = softmax->get_attr<std::string>(op_attr::mode);
                }
            }
        } else {
            mm2 = cur_op;
        }
    }

    VCHECK_SDP_PRIMITIVE(f32_inter, status::invalid_graph,
            "only supports f32 intermediates.");

    auto find_graph_inport = [&inputs](const std::shared_ptr<value_t> &val) {
        auto tmp_val = val;
        while (tmp_val->has_producer()) {
            const op_t &prod_op = tmp_val->get_producer();
            tmp_val = prod_op.get_input_value(0);
        }
        for (int i = 0; i < (int)inputs.size(); i++) {
            if (tmp_val->get_logical_tensor().id == inputs[i].id) { return i; }
        }
        // If the corresponding input is not found, return an invalid value
        return -1;
    };

    VCHECK_SDP_PRIMITIVE(
            mm1 && mm2, status::invalid_graph, "mm1 or mm2 is not found");

    int q_id = find_graph_inport(mm1->get_input_value(0));
    int k_id = find_graph_inport(mm1->get_input_value(1));
    int v_id = find_graph_inport(mm2->get_input_value(1));

    VCHECK_SDP_PRIMITIVE(q_id != -1 && k_id != -1 && v_id != -1,
            status::unimplemented, "Q, K, V are not found");

    VCHECK_SDP_PRIMITIVE(!is_f32 || has_genindex, status::unimplemented,
            "f32 fused sdpa supported for causal mask only");

    // sdp_primitive only supports single scale value.
    if (scale) {
        const auto &s = scale->get_input_logical_tensor(1);
        VCHECK_SDP_PRIMITIVE(ltw(s).nelems() == 1, status::unimplemented,
                "Scale should be single value");
    }

    return status::success;
}

} // namespace dnnl_impl
} // namespace graph
} // namespace impl
} // namespace dnnl