villa/lasagna/grid_sample_3d_u8_diff_kernel.cu at main · ScrollPrize/villa · GitHub

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
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Forward: trilinear sample uint8 volume -> float32 output (no clamping)
__global__ void grid_sample_3d_u8_diff_fwd_kernel(
    const uint8_t* __restrict__ volume,  // (C, Z, Y, X) contiguous
    int C, int Z, int Y, int X,
    const float* __restrict__ grid,      // (D, H, W, 3) possibly non-contiguous
    int N,                               // D*H*W
    long long grid_stride_d,
    long long grid_stride_h,
    long long grid_stride_w,
    long long grid_stride_c,
    int W_grid, int H_grid,
    const float* __restrict__ offset,    // (3,)
    const float* __restrict__ inv_scale, // (3,)
    float* __restrict__ out              // (C, N) contiguous float32
) {
    int n = blockIdx.x * blockDim.x + threadIdx.x;
    if (n >= N) return;

    int w = n % W_grid;
    int tmp = n / W_grid;
    int h = tmp % H_grid;
    int d = tmp / H_grid;

    long long grid_base = d * grid_stride_d + h * grid_stride_h + w * grid_stride_w;
    float px = (grid[grid_base + 0 * grid_stride_c] - offset[0]) * inv_scale[0];
    float py = (grid[grid_base + 1 * grid_stride_c] - offset[1]) * inv_scale[1];
    float pz = (grid[grid_base + 2 * grid_stride_c] - offset[2]) * inv_scale[2];

    float fx0 = floorf(px);
    float fy0 = floorf(py);
    float fz0 = floorf(pz);
    float fx = px - fx0;
    float fy = py - fy0;
    float fz = pz - fz0;

    int ix0 = (int)fx0, iy0 = (int)fy0, iz0 = (int)fz0;
    int ix1 = ix0 + 1, iy1 = iy0 + 1, iz1 = iz0 + 1;

    bool x0_ok = (ix0 >= 0 && ix0 < X);
    bool x1_ok = (ix1 >= 0 && ix1 < X);
    bool y0_ok = (iy0 >= 0 && iy0 < Y);
    bool y1_ok = (iy1 >= 0 && iy1 < Y);
    bool z0_ok = (iz0 >= 0 && iz0 < Z);
    bool z1_ok = (iz1 >= 0 && iz1 < Z);

    float w000 = (1.0f - fx) * (1.0f - fy) * (1.0f - fz);
    float w100 = fx          * (1.0f - fy) * (1.0f - fz);
    float w010 = (1.0f - fx) * fy          * (1.0f - fz);
    float w110 = fx          * fy          * (1.0f - fz);
    float w001 = (1.0f - fx) * (1.0f - fy) * fz;
    float w101 = fx          * (1.0f - fy) * fz;
    float w011 = (1.0f - fx) * fy          * fz;
    float w111 = fx          * fy          * fz;

    long long stride_z = (long long)Y * X;
    long long stride_y = X;

    for (int c = 0; c < C; c++) {
        const uint8_t* vol_c = volume + (long long)c * Z * stride_z;
        float val = 0.0f;

        if (z0_ok && y0_ok && x0_ok) val += w000 * (float)vol_c[(long long)iz0 * stride_z + iy0 * stride_y + ix0];
        if (z0_ok && y0_ok && x1_ok) val += w100 * (float)vol_c[(long long)iz0 * stride_z + iy0 * stride_y + ix1];
        if (z0_ok && y1_ok && x0_ok) val += w010 * (float)vol_c[(long long)iz0 * stride_z + iy1 * stride_y + ix0];
        if (z0_ok && y1_ok && x1_ok) val += w110 * (float)vol_c[(long long)iz0 * stride_z + iy1 * stride_y + ix1];
        if (z1_ok && y0_ok && x0_ok) val += w001 * (float)vol_c[(long long)iz1 * stride_z + iy0 * stride_y + ix0];
        if (z1_ok && y0_ok && x1_ok) val += w101 * (float)vol_c[(long long)iz1 * stride_z + iy0 * stride_y + ix1];
        if (z1_ok && y1_ok && x0_ok) val += w011 * (float)vol_c[(long long)iz1 * stride_z + iy1 * stride_y + ix0];
        if (z1_ok && y1_ok && x1_ok) val += w111 * (float)vol_c[(long long)iz1 * stride_z + iy1 * stride_y + ix1];

        out[(long long)c * N + n] = val;
    }
}


// Backward: compute grad w.r.t. grid positions
__global__ void grid_sample_3d_u8_diff_bwd_kernel(
    const uint8_t* __restrict__ volume,  // (C, Z, Y, X) contiguous
    int C, int Z, int Y, int X,
    const float* __restrict__ grid,      // (D, H, W, 3) possibly non-contiguous
    int N,
    long long grid_stride_d,
    long long grid_stride_h,
    long long grid_stride_w,
    long long grid_stride_c,
    int W_grid, int H_grid,
    const float* __restrict__ offset,    // (3,)
    const float* __restrict__ inv_scale, // (3,)
    const float* __restrict__ grad_output, // (C, N) contiguous
    float* __restrict__ grad_grid          // (N, 3) contiguous
) {
    int n = blockIdx.x * blockDim.x + threadIdx.x;
    if (n >= N) return;

    int w = n % W_grid;
    int tmp = n / W_grid;
    int h = tmp % H_grid;
    int d = tmp / H_grid;

    long long grid_base = d * grid_stride_d + h * grid_stride_h + w * grid_stride_w;
    float isx = inv_scale[0], isy = inv_scale[1], isz = inv_scale[2];
    float px = (grid[grid_base + 0 * grid_stride_c] - offset[0]) * isx;
    float py = (grid[grid_base + 1 * grid_stride_c] - offset[1]) * isy;
    float pz = (grid[grid_base + 2 * grid_stride_c] - offset[2]) * isz;

    float fx0 = floorf(px);
    float fy0 = floorf(py);
    float fz0 = floorf(pz);
    float fx = px - fx0;
    float fy = py - fy0;
    float fz = pz - fz0;

    int ix0 = (int)fx0, iy0 = (int)fy0, iz0 = (int)fz0;
    int ix1 = ix0 + 1, iy1 = iy0 + 1, iz1 = iz0 + 1;

    bool x0_ok = (ix0 >= 0 && ix0 < X);
    bool x1_ok = (ix1 >= 0 && ix1 < X);
    bool y0_ok = (iy0 >= 0 && iy0 < Y);
    bool y1_ok = (iy1 >= 0 && iy1 < Y);
    bool z0_ok = (iz0 >= 0 && iz0 < Z);
    bool z1_ok = (iz1 >= 0 && iz1 < Z);

    long long stride_z = (long long)Y * X;
    long long stride_y = X;

    float dgx = 0.0f, dgy = 0.0f, dgz = 0.0f;

    for (int c = 0; c < C; c++) {
        const uint8_t* vol_c = volume + (long long)c * Z * stride_z;
        float go = grad_output[(long long)c * N + n];

        // Read 8 corner values (0 if out of bounds)
        float v000 = (z0_ok && y0_ok && x0_ok) ? (float)vol_c[(long long)iz0 * stride_z + iy0 * stride_y + ix0] : 0.0f;
        float v100 = (z0_ok && y0_ok && x1_ok) ? (float)vol_c[(long long)iz0 * stride_z + iy0 * stride_y + ix1] : 0.0f;
        float v010 = (z0_ok && y1_ok && x0_ok) ? (float)vol_c[(long long)iz0 * stride_z + iy1 * stride_y + ix0] : 0.0f;
        float v110 = (z0_ok && y1_ok && x1_ok) ? (float)vol_c[(long long)iz0 * stride_z + iy1 * stride_y + ix1] : 0.0f;
        float v001 = (z1_ok && y0_ok && x0_ok) ? (float)vol_c[(long long)iz1 * stride_z + iy0 * stride_y + ix0] : 0.0f;
        float v101 = (z1_ok && y0_ok && x1_ok) ? (float)vol_c[(long long)iz1 * stride_z + iy0 * stride_y + ix1] : 0.0f;
        float v011 = (z1_ok && y1_ok && x0_ok) ? (float)vol_c[(long long)iz1 * stride_z + iy1 * stride_y + ix0] : 0.0f;
        float v111 = (z1_ok && y1_ok && x1_ok) ? (float)vol_c[(long long)iz1 * stride_z + iy1 * stride_y + ix1] : 0.0f;

        // d(interp)/d(fx): differentiate trilinear weights w.r.t. fx
        float dfx = (1.0f - fy) * (1.0f - fz) * (v100 - v000)
                   + fy          * (1.0f - fz) * (v110 - v010)
                   + (1.0f - fy) * fz          * (v101 - v001)
                   + fy          * fz          * (v111 - v011);

        // d(interp)/d(fy)
        float dfy = (1.0f - fx) * (1.0f - fz) * (v010 - v000)
                   + fx          * (1.0f - fz) * (v110 - v100)
                   + (1.0f - fx) * fz          * (v011 - v001)
                   + fx          * fz          * (v111 - v101);

        // d(interp)/d(fz)
        float dfz = (1.0f - fx) * (1.0f - fy) * (v001 - v000)
                   + fx          * (1.0f - fy) * (v101 - v100)
                   + (1.0f - fx) * fy          * (v011 - v010)
                   + fx          * fy          * (v111 - v110);

        // Chain rule: d(grid_coord)/d(fullres_coord) = inv_scale
        dgx += go * dfx * isx;
        dgy += go * dfy * isy;
        dgz += go * dfz * isz;
    }

    grad_grid[n * 3 + 0] = dgx;
    grad_grid[n * 3 + 1] = dgy;
    grad_grid[n * 3 + 2] = dgz;
}


// ---- C++ wrappers ----

torch::Tensor grid_sample_3d_u8_diff_fwd(
    torch::Tensor volume,     // (C, Z, Y, X) uint8 cuda
    torch::Tensor grid,       // (D, H, W, 3) float32 cuda
    torch::Tensor offset,     // (3,) float32 cuda
    torch::Tensor inv_scale   // (3,) float32 cuda
) {
    TORCH_CHECK(volume.is_cuda() && volume.dtype() == torch::kUInt8, "volume must be CUDA uint8");
    TORCH_CHECK(volume.dim() == 4, "volume must be (C, Z, Y, X)");
    TORCH_CHECK(volume.is_contiguous(), "volume must be contiguous");
    TORCH_CHECK(grid.is_cuda() && grid.dtype() == torch::kFloat32, "grid must be CUDA float32");
    TORCH_CHECK(grid.dim() == 4 && grid.size(3) == 3, "grid must be (D, H, W, 3)");
    TORCH_CHECK(offset.is_cuda() && offset.numel() == 3, "offset must be CUDA (3,)");
    TORCH_CHECK(inv_scale.is_cuda() && inv_scale.numel() == 3, "inv_scale must be CUDA (3,)");

    int C = volume.size(0);
    int Z = volume.size(1);
    int Y = volume.size(2);
    int X = volume.size(3);
    int D = grid.size(0);
    int H = grid.size(1);
    int W = grid.size(2);
    int N = D * H * W;

    long long grid_stride_d = grid.stride(0);
    long long grid_stride_h = grid.stride(1);
    long long grid_stride_w = grid.stride(2);
    long long grid_stride_c = grid.stride(3);

    auto out = torch::empty({C, N}, torch::TensorOptions().dtype(torch::kFloat32).device(volume.device()));

    int threads = 256;
    int blocks = (N + threads - 1) / threads;

    grid_sample_3d_u8_diff_fwd_kernel<<<blocks, threads>>>(
        volume.data_ptr<uint8_t>(),
        C, Z, Y, X,
        grid.data_ptr<float>(),
        N,
        grid_stride_d, grid_stride_h, grid_stride_w, grid_stride_c,
        W, H,
        offset.data_ptr<float>(),
        inv_scale.data_ptr<float>(),
        out.data_ptr<float>()
    );

    return out.reshape({C, D, H, W});
}


torch::Tensor grid_sample_3d_u8_diff_bwd(
    torch::Tensor volume,       // (C, Z, Y, X) uint8 cuda
    torch::Tensor grid,         // (D, H, W, 3) float32 cuda
    torch::Tensor offset,       // (3,) float32 cuda
    torch::Tensor inv_scale,    // (3,) float32 cuda
    torch::Tensor grad_output   // (C, D, H, W) float32 cuda
) {
    TORCH_CHECK(volume.is_cuda() && volume.dtype() == torch::kUInt8, "volume must be CUDA uint8");
    TORCH_CHECK(volume.dim() == 4, "volume must be (C, Z, Y, X)");
    TORCH_CHECK(volume.is_contiguous(), "volume must be contiguous");
    TORCH_CHECK(grid.is_cuda() && grid.dtype() == torch::kFloat32, "grid must be CUDA float32");
    TORCH_CHECK(grid.dim() == 4 && grid.size(3) == 3, "grid must be (D, H, W, 3)");
    TORCH_CHECK(grad_output.is_cuda() && grad_output.dtype() == torch::kFloat32, "grad_output must be CUDA float32");
    TORCH_CHECK(grad_output.is_contiguous(), "grad_output must be contiguous");

    int C = volume.size(0);
    int Z = volume.size(1);
    int Y = volume.size(2);
    int X = volume.size(3);
    int D = grid.size(0);
    int H = grid.size(1);
    int W = grid.size(2);
    int N = D * H * W;

    long long grid_stride_d = grid.stride(0);
    long long grid_stride_h = grid.stride(1);
    long long grid_stride_w = grid.stride(2);
    long long grid_stride_c = grid.stride(3);

    auto grad_grid = torch::empty({N, 3}, torch::TensorOptions().dtype(torch::kFloat32).device(volume.device()));

    int threads = 256;
    int blocks = (N + threads - 1) / threads;

    grid_sample_3d_u8_diff_bwd_kernel<<<blocks, threads>>>(
        volume.data_ptr<uint8_t>(),
        C, Z, Y, X,
        grid.data_ptr<float>(),
        N,
        grid_stride_d, grid_stride_h, grid_stride_w, grid_stride_c,
        W, H,
        offset.data_ptr<float>(),
        inv_scale.data_ptr<float>(),
        grad_output.reshape({C, N}).data_ptr<float>(),
        grad_grid.data_ptr<float>()
    );

    return grad_grid.reshape({D, H, W, 3});
}


PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("grid_sample_3d_u8_diff_fwd", &grid_sample_3d_u8_diff_fwd,
          "Trilinear 3D grid_sample on uint8 volume -> float32 (CUDA forward)");
    m.def("grid_sample_3d_u8_diff_bwd", &grid_sample_3d_u8_diff_bwd,
          "Gradient w.r.t. grid positions for trilinear 3D grid_sample (CUDA backward)");
}