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

/*******************************************************************************
* Copyright 2022 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 GEMMSTONE_DSL_IR_CODEGEN_REDUCE_HPP
#define GEMMSTONE_DSL_IR_CODEGEN_REDUCE_HPP

#include "dsl/ir/codegen/register_scope.hpp"
#include "dsl/ir/codegen/reorder.hpp"
#include "dsl/ir/reduce.hpp"
#include "ngen.hpp"

GEMMSTONE_NAMESPACE_START
namespace dsl {
namespace ir {

class reduce_impl_t {
public:
    reduce_impl_t(ngen::HW hw, const reduce_t &reduce, int simd_size)
        : hw_(hw)
        , src_layout_(reduce.src_layout)
        , dst_layout_(reduce.dst_layout)
        , simd_size_(simd_size) {}

    template <typename GeneratorT>
    void emit(GeneratorT *host, ngen_register_scope_t &scope,
            const reg_buf_data_t &src_rd, const reg_buf_data_t &dst_rd) {
        auto &src_type = src_layout_.type();
        auto &dst_type = dst_layout_.type();

        bool is_inplace = (src_rd.base() == dst_rd.base()
                && src_rd.byte_offset() == dst_rd.byte_offset());
        if (is_inplace) {
            dsl_assert(src_type == dst_type)
                    << "Inplace operation is supported for the same type only.";
        }

        auto &back = src_layout_.blocks().back();
        std::vector<bool> seen(back.size * int64_t(back.stride));

        dsl::tile_t tile = find_1d_tile(src_layout_, dst_layout_);
        int tile_elems = (int)tile.elems();
        auto src_tile_layout = src_layout_.sub(tile);
        auto dst_tile_layout = dst_layout_.sub(tile);
        const auto &src_tile_blocks = src_tile_layout.blocks();
        const auto &dst_tile_blocks = dst_tile_layout.blocks();
        dsl_assert(src_tile_blocks.size() <= 1);
        dsl_assert(dst_tile_blocks.size() <= 1);
        ngen_register_scope_t block_scope(scope.register_allocator());
        int src_stride
                = src_tile_blocks.empty() ? 1 : (int)src_tile_blocks[0].stride;
        int dst_stride
                = dst_tile_blocks.empty() ? 1 : (int)dst_tile_blocks[0].stride;
        int grf_size = ngen::GRF::bytes(hw_);
        for (auto &src_start : src_layout_.iter(tile)) {
            ngen_register_scope_t tile_scope(scope.register_allocator());
            auto dst_start = src_start;
            for (size_t i = 0; i < dst_layout_.ndims(); i++) {
                if (dst_layout_.tile()[i] == 1) dst_start[i] = 0;
            }
            int src_off = src_layout_.offset<int>(src_start);
            int dst_off = dst_layout_.offset<int>(dst_start);

            if (is_inplace) {
                bool same_src_dst = (dst_off == src_off);
                if (!seen[dst_off] && !same_src_dst) {
                    dsl_error() << "Invalid inplace reduction.";
                }
                seen[dst_off] = true;
                if (same_src_dst) return;
            }

            auto d = dst_rd.format(dst_off, tile_elems, 1, to_ngen(dst_type));
            auto s = src_rd.format(
                    src_off, tile_elems, src_stride, to_ngen(src_type));
            bool s_half_grf_aligned
                    = one_of(s.byte_offset(), {0, grf_size / 2});
            bool s_is_bf = src_type.is_bf16();
            bool s_is_hf = src_type.is_f16();
            bool s_is_fp8 = src_type.is_fp8();
            bool d_is_f = dst_type.is_f32();
            bool native_bf = host->hw_info().systolic_support();
            bool sd_aligned = (tile_elems == 1
                    || (dst_stride * ngen::getBytes(d.type())
                            == src_stride * ngen::getBytes(s.type())));

            if (src_stride != 1 || !sd_aligned || s_is_hf || s_is_fp8
                    || (s_is_bf && !native_bf)
                    || (s_is_bf && !s_half_grf_aligned)) {
                auto tmp_type = src_type;
                if ((s_is_hf && d_is_f) || s_is_fp8 || (s_is_bf && !native_bf)
                        || ((d.offset() != 0 || !s_half_grf_aligned)
                                && (s_is_bf))) {
                    tmp_type = type_t::f32();
                }
                auto tmp = tile_scope.alloc_reg_data(
                        tmp_type.with_elems(tile_elems));
                emit_reorder_1d_tile(host, tile_scope.register_allocator(),
                        host->hw_info().systolic_support(), tile_elems, s,
                        src_stride, tmp, 1);
                s = tmp.format(0, tile_elems, 1, to_ngen(tmp_type));
            }
            align_src_dst_offset(host, tile_scope, tile_elems, d, s);
            host->add(tile_elems, d.reg_data(), d.reg_data(), s.reg_data());
        }
    }

private:
    dsl::tile_t find_1d_tile(dsl::layout_t a, dsl::layout_t b) const {
        align_layouts(a, b);

        dsl_assert(!a.blocks().empty());
        // Allow trivial tile for scalar dst.
        if (b.blocks().empty()) { return dst_layout_.tile(); }

        auto &a0 = a[0];
        auto &b0 = b[0];

        bool ok = (a0.idx == b0.idx && a0.size == b0.size);
        if (!ok) {
            // Try to match strided layout.
            if (a0.size == 2) {
                auto a_blocks = a.blocks();
                a_blocks.erase(a_blocks.begin());
                a = a.with(a_blocks);
                return find_1d_tile(std::move(a), std::move(b));
            }
            return dsl::tile_t(std::vector<int64_t>(b.ndims(), 1));
        }

        dsl_assert(int64_t(b0.stride) == 1)
                << "Reduction is not supported for non-unit dst stride.";

        int grf_size = ngen::GRF::bytes(hw_);
        int a_grf_elems = grf_size / a.type().size();
        int b_grf_elems = grf_size / b.type().size();

        int min_step = std::min(a_grf_elems, b_grf_elems);
        int max_step = 2 * min_step;

        min_step = std::min(
                std::min(hw_ <= ngen::HW::XeLP ? 8 : simd_size_, min_step),
                (int)a0.size);

        if (a0.size % min_step != 0) {
            // TODO: Extend implementation to support this case.
            stub("Reduction is not supported.");
        }

        std::vector<int64_t> tile_dims(src_layout_.ndims(), 1);
        tile_dims[a0.idx]
                = utils::max_divisor(int(a0.size), {min_step, max_step});

        return dsl::tile_t(tile_dims);
    }

    ngen::HW hw_;
    dsl::layout_t src_layout_;
    dsl::layout_t dst_layout_;
    int simd_size_;
};

} // namespace ir
} // namespace dsl
GEMMSTONE_NAMESPACE_END

#endif