Hugging Face's logo

Spaces:
venite
/
sat3density
Runtime error

App Files Community
sat3density
File size: 13,289 Bytes 
f670afc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
#include <ATen/ATen.h>
#include <ATen/Context.h>
#include <ATen/cuda/CUDAContext.h>

#define CUDA_NUM_THREADS 512 
#define THREADS_PER_BLOCK 64 

#define DIM0(TENSOR) ((TENSOR).x)
#define DIM1(TENSOR) ((TENSOR).y)
#define DIM2(TENSOR) ((TENSOR).z)
#define DIM3(TENSOR) ((TENSOR).w)

#define DIM3_INDEX(TENSOR, xx, yy, zz, ww) ((TENSOR)[((xx) * (TENSOR##_stride.x)) + ((yy) * (TENSOR##_stride.y)) + ((zz) * (TENSOR##_stride.z)) + ((ww) * (TENSOR##_stride.w))])

template <typename scalar_t>
__global__ void kernel_resample2d_update_output(const int n, 
                                               const scalar_t* __restrict__ input1, const long4 input1_size, const long4 input1_stride,
                                               const scalar_t* __restrict__ input2, const long4 input2_size, const long4 input2_stride, 
                                               scalar_t* __restrict__ output,
                                               const long4 output_size, const
                                               long4 output_stride, int
                                               kernel_size/*, bool bilinear*/) {
    int index = blockIdx.x * blockDim.x + threadIdx.x;
    bool bilinear = true;
    if (index >= n) {
        return;
    }

    scalar_t val = 0.0f;

    int dim_b = DIM0(output_size);
    int dim_c = DIM1(output_size);
    int dim_h = DIM2(output_size);
    int dim_w = DIM3(output_size);
    int dim_chw = dim_c * dim_h * dim_w;
    int dim_hw  = dim_h * dim_w;

    int b = ( index / dim_chw ) % dim_b;
    int c = ( index / dim_hw )  % dim_c;
    int y = ( index / dim_w )   % dim_h;
    int x = ( index          )  % dim_w;

    scalar_t dx = DIM3_INDEX(input2, b, 0, y, x);
    scalar_t dy = DIM3_INDEX(input2, b, 1, y, x);

    scalar_t xf = static_cast<scalar_t>(x) + dx;
    scalar_t yf = static_cast<scalar_t>(y) + dy;
    scalar_t alpha = xf - floor(xf); // alpha
    scalar_t beta = yf - floor(yf); // beta

    if (bilinear) {
        int xL = max(min( int (floor(xf)),    dim_w-1), 0);
        int xR = max(min( int (floor(xf)+1), dim_w -1), 0);
        int yT = max(min( int (floor(yf)),    dim_h-1), 0);
        int yB = max(min( int (floor(yf)+1),  dim_h-1), 0);

        for (int fy = 0; fy < kernel_size; fy += 1) {
            for (int fx = 0; fx < kernel_size; fx += 1) {
                val += static_cast<float>((1. - alpha)*(1. - beta) * DIM3_INDEX(input1, b, c, yT + fy, xL + fx));
                val += static_cast<float>((alpha)*(1. - beta) * DIM3_INDEX(input1, b, c, yT + fy, xR + fx));
                val += static_cast<float>((1. - alpha)*(beta) * DIM3_INDEX(input1, b, c, yB + fy, xL + fx));
                val += static_cast<float>((alpha)*(beta) * DIM3_INDEX(input1, b, c, yB + fy, xR + fx));
            }
        }

        output[index] = val;
    }
    else {
        int xN = max(min( int (floor(xf + 0.5)), dim_w - 1), 0);
        int yN = max(min( int (floor(yf + 0.5)), dim_h - 1), 0);

        output[index] = static_cast<float> ( DIM3_INDEX(input1, b, c, yN, xN) );
    }

}


template <typename scalar_t>
__global__ void kernel_resample2d_backward_input1(
    const int n, const scalar_t* __restrict__ input1, const long4 input1_size, const long4 input1_stride,
    const scalar_t* __restrict__ input2, const long4 input2_size, const long4 input2_stride,
    const scalar_t* __restrict__ gradOutput, const long4 gradOutput_size, const long4 gradOutput_stride,
    scalar_t* __restrict__ gradInput, const long4 gradInput_size, const long4
    gradInput_stride, int kernel_size/*, bool bilinear*/) {

    int index = blockIdx.x * blockDim.x + threadIdx.x;
    bool bilinear = true;
    if (index >= n) {
        return;
    }

    int dim_b = DIM0(gradOutput_size);
    int dim_c = DIM1(gradOutput_size);
    int dim_h = DIM2(gradOutput_size);
    int dim_w = DIM3(gradOutput_size);
    int dim_chw = dim_c * dim_h * dim_w;
    int dim_hw  = dim_h * dim_w;

    int b = ( index / dim_chw ) % dim_b;
    int c = ( index / dim_hw )  % dim_c;
    int y = ( index / dim_w )   % dim_h;
    int x = ( index          )  % dim_w;

    scalar_t dx = DIM3_INDEX(input2, b, 0, y, x);
    scalar_t dy = DIM3_INDEX(input2, b, 1, y, x);

    scalar_t xf = static_cast<scalar_t>(x) + dx;
    scalar_t yf = static_cast<scalar_t>(y) + dy;
    scalar_t alpha = xf - int(xf); // alpha
    scalar_t beta = yf - int(yf); // beta

    int idim_h = DIM2(input1_size);
    int idim_w = DIM3(input1_size);

    int xL = max(min( int (floor(xf)),    idim_w-1), 0);
    int xR = max(min( int (floor(xf)+1), idim_w -1), 0);
    int yT = max(min( int (floor(yf)),    idim_h-1), 0);
    int yB = max(min( int (floor(yf)+1),  idim_h-1), 0);

    for (int fy = 0; fy < kernel_size; fy += 1) {
        for (int fx = 0; fx < kernel_size; fx += 1) {
            atomicAdd(&DIM3_INDEX(gradInput, b, c, (yT + fy), (xL + fx)), (1-alpha)*(1-beta) * DIM3_INDEX(gradOutput, b, c, y, x));
            atomicAdd(&DIM3_INDEX(gradInput, b, c, (yT + fy), (xR + fx)),   (alpha)*(1-beta) * DIM3_INDEX(gradOutput, b, c, y, x));
            atomicAdd(&DIM3_INDEX(gradInput, b, c, (yB + fy), (xL + fx)),   (1-alpha)*(beta) * DIM3_INDEX(gradOutput, b, c, y, x));
            atomicAdd(&DIM3_INDEX(gradInput, b, c, (yB + fy), (xR + fx)),     (alpha)*(beta) * DIM3_INDEX(gradOutput, b, c, y, x));
        }
    }

}

template <typename scalar_t>
__global__ void kernel_resample2d_backward_input2(
    const int n, const scalar_t* __restrict__ input1, const long4 input1_size, const long4 input1_stride,
    const scalar_t* __restrict__ input2, const long4 input2_size, const long4 input2_stride,
    const scalar_t* __restrict__ gradOutput, const long4 gradOutput_size, const long4 gradOutput_stride,
    scalar_t* __restrict__ gradInput, const long4 gradInput_size, const long4
    gradInput_stride, int kernel_size/*, bool bilinear*/) {

    int index = blockIdx.x * blockDim.x + threadIdx.x;
    bool bilinear = true;
    if (index >= n) {
        return;
    }

    scalar_t output = 0.0;
    int kernel_rad = (kernel_size - 1)/2;

    int dim_b = DIM0(gradInput_size);
    int dim_c = DIM1(gradInput_size);
    int dim_h = DIM2(gradInput_size);
    int dim_w = DIM3(gradInput_size);
    int dim_chw = dim_c * dim_h * dim_w;
    int dim_hw  = dim_h * dim_w;

    int b = ( index / dim_chw ) % dim_b;
    int c = ( index / dim_hw )  % dim_c;
    int y = ( index / dim_w )   % dim_h;
    int x = ( index          )  % dim_w;

    int odim_c = DIM1(gradOutput_size);

    scalar_t dx = DIM3_INDEX(input2, b, 0, y, x);
    scalar_t dy = DIM3_INDEX(input2, b, 1, y, x);

    scalar_t xf = static_cast<scalar_t>(x) + dx;
    scalar_t yf = static_cast<scalar_t>(y) + dy;

    int xL = max(min( int (floor(xf)),    dim_w-1), 0);
    int xR = max(min( int (floor(xf)+1), dim_w -1), 0);
    int yT = max(min( int (floor(yf)),    dim_h-1), 0);
    int yB = max(min( int (floor(yf)+1),  dim_h-1), 0);
    
    if (c % 2) {
        float gamma = 1 - (xf - floor(xf)); // alpha
        for (int i = 0; i <= 2*kernel_rad; ++i) {
            for (int j = 0; j <= 2*kernel_rad; ++j) {
                for (int ch = 0; ch < odim_c; ++ch) {
                    output += (gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yB + j), (xL + i));
                    output -= (gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yT + j), (xL + i));
                    output += (1-gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yB + j), (xR + i));
                    output -= (1-gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yT + j), (xR + i));
                }
            }
        }
    }
    else {
        float gamma = 1 - (yf - floor(yf)); // alpha
        for (int i = 0; i <= 2*kernel_rad; ++i) {
            for (int j = 0; j <= 2*kernel_rad; ++j) {
                for (int ch = 0; ch < odim_c; ++ch) {
                    output += (gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yT + j), (xR + i));
                    output -= (gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yT + j), (xL + i));
                    output += (1-gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yB + j), (xR + i));
                    output -= (1-gamma) * DIM3_INDEX(gradOutput, b, ch, y, x) * DIM3_INDEX(input1, b, ch, (yB + j), (xL + i));
                }
            }
        }

    }

    gradInput[index] = output;

}

void resample2d_kernel_forward(
    at::Tensor& input1, 
    at::Tensor& input2,
    at::Tensor& output, 
    int kernel_size/*,
    bool bilinear*/) {

    int n = output.numel();

    const long4 input1_size = make_long4(input1.size(0), input1.size(1), input1.size(2), input1.size(3));
    const long4 input1_stride = make_long4(input1.stride(0), input1.stride(1), input1.stride(2), input1.stride(3));

    const long4 input2_size = make_long4(input2.size(0), input2.size(1), input2.size(2), input2.size(3));
    const long4 input2_stride = make_long4(input2.stride(0), input2.stride(1), input2.stride(2), input2.stride(3));

    const long4 output_size = make_long4(output.size(0), output.size(1), output.size(2), output.size(3));
    const long4 output_stride = make_long4(output.stride(0), output.stride(1), output.stride(2), output.stride(3));

    // TODO: when atomicAdd gets resolved, change to AT_DISPATCH_FLOATING_TYPES_AND_HALF
//    AT_DISPATCH_FLOATING_TYPES(input1.type(), "resample_forward_kernel", ([&] {

        kernel_resample2d_update_output<float><<< (n + CUDA_NUM_THREADS - 1)/CUDA_NUM_THREADS, CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream() >>>(
//at::globalContext().getCurrentCUDAStream() >>>(
            n,
            input1.data<float>(),
            input1_size,
            input1_stride, 
            input2.data<float>(),
            input2_size,
            input2_stride,
            output.data<float>(),
            output_size,
            output_stride,
            kernel_size/*,
            bilinear*/);

//    }));

        // TODO: ATen-equivalent check

       //    THCudaCheck(cudaGetLastError());

}

void resample2d_kernel_backward(
    at::Tensor& input1,
    at::Tensor& input2,
    at::Tensor& gradOutput,
    at::Tensor& gradInput1,
    at::Tensor& gradInput2,
    int kernel_size/*,
    bool bilinear*/) {

    int n = gradOutput.numel();

    const long4 input1_size = make_long4(input1.size(0), input1.size(1), input1.size(2), input1.size(3));
    const long4 input1_stride = make_long4(input1.stride(0), input1.stride(1), input1.stride(2), input1.stride(3));

    const long4 input2_size = make_long4(input2.size(0), input2.size(1), input2.size(2), input2.size(3));
    const long4 input2_stride = make_long4(input2.stride(0), input2.stride(1), input2.stride(2), input2.stride(3));

    const long4 gradOutput_size = make_long4(gradOutput.size(0), gradOutput.size(1), gradOutput.size(2), gradOutput.size(3));
    const long4 gradOutput_stride = make_long4(gradOutput.stride(0), gradOutput.stride(1), gradOutput.stride(2), gradOutput.stride(3));

    const long4 gradInput1_size = make_long4(gradInput1.size(0), gradInput1.size(1), gradInput1.size(2), gradInput1.size(3));
    const long4 gradInput1_stride = make_long4(gradInput1.stride(0), gradInput1.stride(1), gradInput1.stride(2), gradInput1.stride(3));

//    AT_DISPATCH_FLOATING_TYPES(input1.type(), "resample_backward_input1", ([&] {

        kernel_resample2d_backward_input1<float><<< (n + CUDA_NUM_THREADS - 1)/CUDA_NUM_THREADS, CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream() >>>(
//at::globalContext().getCurrentCUDAStream() >>>(
            n, 
            input1.data<float>(), 
            input1_size,
            input1_stride,
            input2.data<float>(),
            input2_size, 
            input2_stride,
            gradOutput.data<float>(),
            gradOutput_size,
            gradOutput_stride,
            gradInput1.data<float>(),
            gradInput1_size,
            gradInput1_stride, 
            kernel_size/*,
            bilinear*/
        );

//    }));

    const long4 gradInput2_size = make_long4(gradInput2.size(0), gradInput2.size(1), gradInput2.size(2), gradInput2.size(3));
    const long4 gradInput2_stride = make_long4(gradInput2.stride(0), gradInput2.stride(1), gradInput2.stride(2), gradInput2.stride(3));

    n = gradInput2.numel();

//    AT_DISPATCH_FLOATING_TYPES(gradInput2.type(), "resample_backward_input2", ([&] {


        kernel_resample2d_backward_input2<float><<< (n + CUDA_NUM_THREADS - 1)/CUDA_NUM_THREADS, CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream() >>>(
//at::globalContext().getCurrentCUDAStream() >>>(
            n, 
            input1.data<float>(), 
            input1_size, 
            input1_stride,
            input2.data<float>(), 
            input2_size,
            input2_stride,
            gradOutput.data<float>(),
            gradOutput_size,
            gradOutput_stride,
            gradInput2.data<float>(),
            gradInput2_size,
            gradInput2_stride,
            kernel_size/*,
            bilinear*/
       );

//    }));

    // TODO: Use the ATen equivalent to get last error

    //    THCudaCheck(cudaGetLastError());

}