Hugging Face's logo

Spaces:
qbhf2
/
GarmentCode
Sleeping

App Files Community
GarmentCode
File size: 9,615 Bytes
66c9c8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
 
/** Copyright (c) 2022 NVIDIA CORPORATION.  All rights reserved.
 * NVIDIA CORPORATION and its licensors retain all intellectual property
 * and proprietary rights in and to this software, related documentation
 * and any modifications thereto.  Any use, reproduction, disclosure or
 * distribution of this software and related documentation without an express
 * license agreement from NVIDIA CORPORATION is strictly prohibited.
 */

#pragma once

namespace wp
{


CUDA_CALLABLE inline int dense_index(int stride, int i, int j)
{
    return i*stride + j;
}

template <bool transpose>
CUDA_CALLABLE inline int dense_index(int rows, int cols, int i, int j)
{
    if (transpose)
        return j*rows + i;
    else
        return i*cols + j;
}



template <bool t1, bool t2, bool add>
CUDA_CALLABLE inline void dense_gemm_impl(int m, int n, int p, const float* __restrict__ A, const float* __restrict__ B, float* __restrict__ C)
{
    for (int i=0; i < m; i++)
    {
        for (int j=0; j < n; ++j)
        {
            float sum = 0.0f;

            for (int k=0; k < p; ++k)
            {
                sum += A[dense_index<t1>(m, p, i, k)]*B[dense_index<t2>(p, n, k, j)];
            }
            
            if (add)
                C[i*n + j] += sum;
            else
                C[i*n + j] = sum;
        }
    }
}


template <bool add=false>
CUDA_CALLABLE inline void dense_gemm(int m, int n, int p, int t1, int t2, const array_t<float>& A, const array_t<float>& B, array_t<float>& C)
{
    if (t1 == 0 && t2 == 0)
        dense_gemm_impl<false, false, add>(m, n, p, A.data, B.data, C.data);
    else if (t1 == 1 && t2 == 0)
        dense_gemm_impl<true, false, add>(m, n, p, A.data, B.data, C.data);
    else if (t1 == 0 && t2 == 1)
        dense_gemm_impl<false, true, add>(m, n, p, A.data, B.data, C.data);
    else if (t1 == 1 && t2 == 1)
        dense_gemm_impl<true, true, add>(m, n, p, A.data, B.data, C.data);
}




void  CUDA_CALLABLE inline dense_chol(int n, const array_t<float>& A, float regularization, array_t<float>& L)
{
    for (int j=0; j < n; ++j)
    {
        float s = A.data[dense_index(n, j, j)] + regularization;

        for (int k=0; k < j; ++k)
        {
            float r = L.data[dense_index(n, j, k)];
            s -= r*r;
        }

        s = sqrt(s);
        const float invS = 1.0f/s;

        L.data[dense_index(n, j, j)] = s;

        for (int i=j+1; i < n; ++i)
        {
            s = A.data[dense_index(n, i, j)];
            
            for (int k=0; k < j; ++k)
            {
                s -= L.data[dense_index(n, i, k)]*L.data[dense_index(n, j, k)];
            }

            L.data[dense_index(n, i, j)] = s*invS;
        }
    }
}




// Solves (L*L^T)x = b given the Cholesky factor L 
CUDA_CALLABLE inline void dense_subs(int n, const array_t<float>& L, const array_t<float>& b, array_t<float>& x)
{
    // forward substitution
    for (int i=0; i < n; ++i)
    {
        float s = b.data[i];

        for (int j=0; j < i; ++j)
        {
            s -= L.data[dense_index(n, i, j)]*x.data[j];
        }

        x.data[i] = s/L.data[dense_index(n, i, i)];
    }

    // backward substitution
    for (int i=n-1; i >= 0; --i)
    {
        float s = x.data[i];

        for (int j=i+1; j < n; ++j)
        {
            s -= L.data[dense_index(n, j, i)]*x.data[j];
        }

        x.data[i] = s/L.data[dense_index(n, i, i)];
    }
}

CUDA_CALLABLE inline void dense_solve(int n, const array_t<float>& A, const array_t<float>& L, const array_t<float>& b, array_t<float>& x)
{
    dense_subs(n, L, b, x);
}


// CUDA_CALLABLE inline void print_matrix(const char* name, int m, int n, const float* data)
// {
//     printf("%s = [", name);

//     for (int i=0; i < m; ++i)
//     {
//         for (int j=0; j < n; ++j)
//         {
//             printf("%f ", data[dense_index(n, i, j)]);
//         }

//         printf(";\n");
//     }

//     printf("]\n");
// }

// adjoint methods
CUDA_CALLABLE inline void adj_dense_gemm(
    int m, int n, int p, int t1, int t2, const array_t<float>& A, const array_t<float>& B, array_t<float>& C,
    int adj_m, int adj_n, int adj_p, int adj_t1, int adj_t2, array_t<float>& adj_A, array_t<float>& adj_B, const array_t<float>& adj_C)
{

    // print_matrix("A", m, p, A);
    // print_matrix("B", p, n, B);
    // printf("t1: %d t2: %d\n", t1, t2);

    if (t1)
    {
        dense_gemm<true>(p, m, n, 0, 1, B, adj_C, adj_A);
        dense_gemm<true>(p, n, m, int(!t1), 0, A, adj_C, adj_B);
    }
    else
    {
        dense_gemm<true>(m, p, n, 0, int(!t2), adj_C, B, adj_A);
        dense_gemm<true>(p, n, m, int(!t1), 0, A, adj_C, adj_B);
    }
}


CUDA_CALLABLE inline void adj_dense_chol(
    int n, const array_t<float>& A, float regularization, array_t<float>& L,
    int adj_n, const array_t<float>& adj_A, float adj_regularization, array_t<float>& adj_L)
{
    // nop, use dense_solve to differentiate through (A^-1)b = x
}

CUDA_CALLABLE inline void adj_dense_subs(
    int n, const array_t<float>& L, const array_t<float>& b, array_t<float>& x,
    int adj_n, const array_t<float>& adj_L, const array_t<float>& adj_b, array_t<float>& adj_x)
{
    // nop, use dense_solve to differentiate through (A^-1)b = x
}


CUDA_CALLABLE inline void adj_dense_solve(int n,
    const array_t<float>& A, const array_t<float>& L, const array_t<float>& b, const array_t<float>& x,
    int adj_n, array_t<float>& adj_A, array_t<float>& adj_L, array_t<float>& adj_b, const array_t<float>& adj_x)
{
    // see https://people.maths.ox.ac.uk/gilesm/files/NA-08-01.pwp, section 2.3.1
    dense_subs(n, L, adj_x, adj_b);

    // A* = -adj_b*x^T
    for (int i=0; i < n; ++i)
    {
        for (int j=0; j < n; ++j)
        {
            adj_A.data[dense_index(n, i, j)] += -adj_b.data[i]*x.data[j];
        }
    }
}


template <typename F>
CUDA_CALLABLE inline void mlp(const array_t<float>& weights, const array_t<float>& bias, F activation, int index, const array_t<float>& x, array_t<float>& out)
{
    const int m = weights.shape[0];
    const int n = weights.shape[1];
    const int b = x.shape[1];

    for (int i=0; i < m; ++i)
    {
        float tmp = bias.data[i];

        for(int j=0; j < n; ++j)
        {
            tmp += weights.data[i*n + j]*x.data[index + b*j];
        }

        out.data[index + b*i] = activation(tmp);
    }
}

template <typename F, typename AdjF>
CUDA_CALLABLE inline void adj_mlp(const array_t<float>& weights, const array_t<float>& bias, F activation, int index, const array_t<float>& x, array_t<float>& out,
                                  array_t<float>& adj_weights, array_t<float>& adj_bias, AdjF adj_activation, int adj_index, array_t<float>& adj_x, array_t<float>& adj_out)
{
    const int m = weights.shape[0];
    const int n = weights.shape[1];
    const int b = x.shape[1];

    for (int i=0; i < m; ++i)
    {
        // recompute forward pass so we don't have to store pre-activation outputs
        float tmp = bias.data[i];

        for(int j=0; j < n; ++j)
        {
            tmp += weights.data[i*n + j]*x.data[index + b*j];
        }

        // adjoint w.r.t to acivation
        float adj_f = 0.0f;
    
        if (adj_out.data)
            adj_activation(tmp, adj_f, adj_out.data[index + b*i]);

        for (int j=0; j < n; ++j)
        {
            // adjoint w.r.t M_i
            if (adj_weights.data)
                atomic_add(&adj_weights.data[i*n + j], x.data[index + b*j]*adj_f);    // todo: reduce these atomic stores using warp/block level reductions

            // adjoint w.r.t x
            if (adj_x.data)
                atomic_add(&adj_x.data[index + b*j], weights.data[i*n + j]*adj_f);
        }

        // adjoint w.r.t b
        if (adj_bias.data)
            atomic_add(&adj_bias.data[i], adj_f);

    }
}


// template <typename F>
// CUDA_CALLABLE inline void mlp(const array_t<float>& weights, const array_t<float>& bias, F activation, int m, int n, int b, int index, const array_t<float>& x, array_t<float>& out)
// {
//     x += index*n;
//     out += index*m;


//     for (int i=0; i < m; ++i)
//     {
//         float tmp = bias[i];

//         for(int j=0; j < n; ++j)
//         {
//             tmp += weights[i*n + j]*x[j];
//         }

//         out[i] = activation(tmp);
//     }
// }

// template <typename F, typename AdjF>
// CUDA_CALLABLE inline void adj_mlp(const array_t<float>& weights, const array_t<float>& bias, F activation, int m, int n, int b, int index, const array_t<float>& x, const array_t<float>& out,
//                                   array_t<float>& adj_weights, array_t<float>& adj_bias, AdjF adj_activation, int adj_m, int adj_n, int adj_b, int adj_index, array_t<float>& adj_x, array_t<float>& adj_out)
// {
//     x += index*n;
//     out += index*m;

//     adj_x += index*n;
//     adj_out += index*m;

//     for (int i=0; i < m; ++i)
//     {
//         // recompute forward pass so we don't have to store pre-activation outputs
//         float tmp = bias[i];

//         for(int j=0; j < n; ++j)
//         {
//             tmp += weights[i*n + j]*x[index + b*j];            
//         }

//         // adjoint w.r.t to acivation
//         float adj_f = 0.0f;
//         adj_activation(tmp, adj_f, adj_out[index + b*i]);

//         for (int j=0; j < n; ++j)
//         {
//             // adjoint w.r.t M_i
//             adj_weights[i*n + j] += x[j]*adj_f;

//             // adjoint w.r.t x
//             adj_x[index + b*j] += weights[i*n + j]*adj_f;
//         }

//         // adjoint w.r.t b
//         adj_bias[i] += adj_f;
//     }
// }

} // namespace wp