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
240
241
242
243

/*******************************************************************************
* Copyright 2025 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 <math.h>
#include <riscv_vector.h>

#include "common/dnnl_thread.hpp"
#include "common/memory_desc_wrapper.hpp"

#include "cpu/rv64/rvv_group_normalization.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace rv64 {
namespace {

inline void stats_reduction(
        const float *src, size_t len, double &sum_out, double &sumsq_out) {
    size_t vl_max = __riscv_vsetvlmax_e32m1();

    vfloat64m2_t v_sum0 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sum1 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sum2 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sum3 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);

    vfloat64m2_t v_sq0 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sq1 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sq2 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);
    vfloat64m2_t v_sq3 = __riscv_vfmv_v_f_f64m2(0.0, vl_max);

    size_t idx = 0;

    for (; idx + 4 * vl_max <= len; idx += 4 * vl_max) {
        vfloat32m1_t v_x0
                = __riscv_vle32_v_f32m1(src + idx + 0 * vl_max, vl_max);
        vfloat32m1_t v_x1
                = __riscv_vle32_v_f32m1(src + idx + 1 * vl_max, vl_max);
        vfloat32m1_t v_x2
                = __riscv_vle32_v_f32m1(src + idx + 2 * vl_max, vl_max);
        vfloat32m1_t v_x3
                = __riscv_vle32_v_f32m1(src + idx + 3 * vl_max, vl_max);

        v_sum0 = __riscv_vfwadd_wv_f64m2(v_sum0, v_x0, vl_max);
        v_sum1 = __riscv_vfwadd_wv_f64m2(v_sum1, v_x1, vl_max);
        v_sum2 = __riscv_vfwadd_wv_f64m2(v_sum2, v_x2, vl_max);
        v_sum3 = __riscv_vfwadd_wv_f64m2(v_sum3, v_x3, vl_max);

        v_sq0 = __riscv_vfwmacc_vv_f64m2(v_sq0, v_x0, v_x0, vl_max);
        v_sq1 = __riscv_vfwmacc_vv_f64m2(v_sq1, v_x1, v_x1, vl_max);
        v_sq2 = __riscv_vfwmacc_vv_f64m2(v_sq2, v_x2, v_x2, vl_max);
        v_sq3 = __riscv_vfwmacc_vv_f64m2(v_sq3, v_x3, v_x3, vl_max);
    }

    vfloat64m2_t v_sum_all = __riscv_vfadd_vv_f64m2(
            __riscv_vfadd_vv_f64m2(v_sum0, v_sum1, vl_max),
            __riscv_vfadd_vv_f64m2(v_sum2, v_sum3, vl_max), vl_max);
    vfloat64m2_t v_sq_all = __riscv_vfadd_vv_f64m2(
            __riscv_vfadd_vv_f64m2(v_sq0, v_sq1, vl_max),
            __riscv_vfadd_vv_f64m2(v_sq2, v_sq3, vl_max), vl_max);

    while (idx < len) {
        size_t vl = __riscv_vsetvl_e32m1(len - idx);
        vfloat32m1_t v_x = __riscv_vle32_v_f32m1(src + idx, vl);
        v_sum_all = __riscv_vfwadd_wv_f64m2(v_sum_all, v_x, vl);
        v_sq_all = __riscv_vfwmacc_vv_f64m2(v_sq_all, v_x, v_x, vl);
        idx += vl;
    }

    vfloat64m1_t v_red_zero = __riscv_vfmv_v_f_f64m1(0.0, vl_max);
    sum_out = __riscv_vfmv_f_s_f64m1_f64(
            __riscv_vfredusum_vs_f64m2_f64m1(v_sum_all, v_red_zero, vl_max));
    sumsq_out = __riscv_vfmv_f_s_f64m1_f64(
            __riscv_vfredusum_vs_f64m2_f64m1(v_sq_all, v_red_zero, vl_max));
}

inline void norm_spatial_loop(const float *src, float *dst, size_t len,
        float mean_val, float inv_std_val, float gamma_val, float beta_val,
        bool use_scale, bool use_shift) {

    size_t vl_max = __riscv_vsetvlmax_e32m1();
    size_t idx = 0;

    vfloat32m1_t v_mean = __riscv_vfmv_v_f_f32m1(mean_val, vl_max);
    vfloat32m1_t v_inv_std = __riscv_vfmv_v_f_f32m1(inv_std_val, vl_max);
    vfloat32m1_t v_gamma = __riscv_vfmv_v_f_f32m1(gamma_val, vl_max);
    vfloat32m1_t v_beta = __riscv_vfmv_v_f_f32m1(beta_val, vl_max);

    for (; idx + 4 * vl_max <= len; idx += 4 * vl_max) {
        vfloat32m1_t v0 = __riscv_vle32_v_f32m1(src + idx + 0 * vl_max, vl_max);
        vfloat32m1_t v1 = __riscv_vle32_v_f32m1(src + idx + 1 * vl_max, vl_max);
        vfloat32m1_t v2 = __riscv_vle32_v_f32m1(src + idx + 2 * vl_max, vl_max);
        vfloat32m1_t v3 = __riscv_vle32_v_f32m1(src + idx + 3 * vl_max, vl_max);

        v0 = __riscv_vfsub_vv_f32m1(v0, v_mean, vl_max);
        v1 = __riscv_vfsub_vv_f32m1(v1, v_mean, vl_max);
        v2 = __riscv_vfsub_vv_f32m1(v2, v_mean, vl_max);
        v3 = __riscv_vfsub_vv_f32m1(v3, v_mean, vl_max);

        v0 = __riscv_vfmul_vv_f32m1(v0, v_inv_std, vl_max);
        v1 = __riscv_vfmul_vv_f32m1(v1, v_inv_std, vl_max);
        v2 = __riscv_vfmul_vv_f32m1(v2, v_inv_std, vl_max);
        v3 = __riscv_vfmul_vv_f32m1(v3, v_inv_std, vl_max);

        if (use_scale) {
            v0 = __riscv_vfmul_vv_f32m1(v0, v_gamma, vl_max);
            v1 = __riscv_vfmul_vv_f32m1(v1, v_gamma, vl_max);
            v2 = __riscv_vfmul_vv_f32m1(v2, v_gamma, vl_max);
            v3 = __riscv_vfmul_vv_f32m1(v3, v_gamma, vl_max);
        }
        if (use_shift) {
            v0 = __riscv_vfadd_vv_f32m1(v0, v_beta, vl_max);
            v1 = __riscv_vfadd_vv_f32m1(v1, v_beta, vl_max);
            v2 = __riscv_vfadd_vv_f32m1(v2, v_beta, vl_max);
            v3 = __riscv_vfadd_vv_f32m1(v3, v_beta, vl_max);
        }

        __riscv_vse32_v_f32m1(dst + idx + 0 * vl_max, v0, vl_max);
        __riscv_vse32_v_f32m1(dst + idx + 1 * vl_max, v1, vl_max);
        __riscv_vse32_v_f32m1(dst + idx + 2 * vl_max, v2, vl_max);
        __riscv_vse32_v_f32m1(dst + idx + 3 * vl_max, v3, vl_max);
    }

    while (idx < len) {
        size_t vl = __riscv_vsetvl_e32m1(len - idx);
        vfloat32m1_t v_x = __riscv_vle32_v_f32m1(src + idx, vl);

        v_x = __riscv_vfsub_vf_f32m1(v_x, mean_val, vl);
        v_x = __riscv_vfmul_vf_f32m1(v_x, inv_std_val, vl);

        if (use_scale) v_x = __riscv_vfmul_vf_f32m1(v_x, gamma_val, vl);
        if (use_shift) v_x = __riscv_vfadd_vf_f32m1(v_x, beta_val, vl);

        __riscv_vse32_v_f32m1(dst + idx, v_x, vl);
        idx += vl;
    }
}

} // namespace

status_t rvv_group_normalization_fwd_t::execute_forward(
        const exec_ctx_t &ctx) const {
    const memory_desc_wrapper src_d(pd()->src_md());

    auto src = CTX_IN_MEM(const float *, DNNL_ARG_SRC);
    auto dst = CTX_OUT_MEM(float *, DNNL_ARG_DST);
    const float *scale = pd()->use_scale()
            ? CTX_IN_MEM(const float *, DNNL_ARG_SCALE)
            : nullptr;
    const float *shift = pd()->use_shift()
            ? CTX_IN_MEM(const float *, DNNL_ARG_SHIFT)
            : nullptr;

    float *mean = pd()->stats_is_src()
            ? const_cast<float *>(CTX_IN_MEM(const float *, DNNL_ARG_MEAN))
            : CTX_OUT_MEM(float *, DNNL_ARG_MEAN);
    float *variance = pd()->stats_is_src()
            ? const_cast<float *>(CTX_IN_MEM(const float *, DNNL_ARG_VARIANCE))
            : CTX_OUT_MEM(float *, DNNL_ARG_VARIANCE);

    const auto N = pd()->MB();
    const auto C = pd()->C();
    const auto D = pd()->D();
    const auto H = pd()->H();
    const auto W = pd()->W();
    const size_t SP = D * H * W;

    const auto G = pd()->desc()->groups;
    const auto eps = pd()->desc()->group_norm_epsilon;
    const auto calculate_stats = !pd()->stats_is_src();
    const auto save_stats = pd()->is_training();

    const auto C_PER_G = C / G;

    parallel_nd(N, G, [&](dim_t n, dim_t g) {
        dim_t c_start = g * C_PER_G;
        size_t group_off = n * C * SP + c_start * SP;
        size_t group_len = C_PER_G * SP;

        float v_mean = 0.0f;
        float v_var = 0.0f;

        if (calculate_stats) {
            double sum = 0.0;
            double sumsq = 0.0;
            stats_reduction(src + group_off, group_len, sum, sumsq);

            double mean_d = sum / (double)group_len;
            double var_d = sumsq / (double)group_len - mean_d * mean_d;

            if (var_d < 0) var_d = 0;

            v_mean = (float)mean_d;
            v_var = (float)var_d;

            if (save_stats) {
                size_t stat_off = n * G + g;
                mean[stat_off] = v_mean;
                variance[stat_off] = v_var;
            }
        } else {
            size_t stat_off = n * G + g;
            v_mean = mean[stat_off];
            v_var = variance[stat_off];
        }

        float inv_std = 1.0f / sqrtf(v_var + eps);

        for (dim_t c = 0; c < C_PER_G; ++c) {
            dim_t global_c = c_start + c;
            size_t channel_off = group_off + c * SP;

            float gamma = scale ? scale[global_c] : 1.0f;
            float beta = shift ? shift[global_c] : 0.0f;

            bool use_scale = (scale != nullptr);
            bool use_shift = (shift != nullptr);

            norm_spatial_loop(src + channel_off, dst + channel_off, SP, v_mean,
                    inv_std, gamma, beta, use_scale, use_shift);
        }
    });

    return status::success;
}

} // namespace rv64
} // namespace cpu
} // namespace impl
} // namespace dnnl