Spaces:
Running
Running
File size: 7,808 Bytes
f5ba9ea |
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 |
// Copyright (c) Meta Platforms, Inc. and affiliates.
// All rights reserved.
// This source code is licensed under the license found in the
// LICENSE file in the root directory of this source tree.
// adapted from https://github.com/zsef123/Connected_components_PyTorch
// with license found in the LICENSE_cctorch file in the root directory.
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <torch/extension.h>
#include <torch/script.h>
#include <vector>
// 2d
#define BLOCK_ROWS 16
#define BLOCK_COLS 16
namespace cc2d {
template <typename T>
__device__ __forceinline__ unsigned char hasBit(T bitmap, unsigned char pos) {
return (bitmap >> pos) & 1;
}
__device__ int32_t find(const int32_t* s_buf, int32_t n) {
while (s_buf[n] != n)
n = s_buf[n];
return n;
}
__device__ int32_t find_n_compress(int32_t* s_buf, int32_t n) {
const int32_t id = n;
while (s_buf[n] != n) {
n = s_buf[n];
s_buf[id] = n;
}
return n;
}
__device__ void union_(int32_t* s_buf, int32_t a, int32_t b) {
bool done;
do {
a = find(s_buf, a);
b = find(s_buf, b);
if (a < b) {
int32_t old = atomicMin(s_buf + b, a);
done = (old == b);
b = old;
} else if (b < a) {
int32_t old = atomicMin(s_buf + a, b);
done = (old == a);
a = old;
} else
done = true;
} while (!done);
}
__global__ void
init_labeling(int32_t* label, const uint32_t W, const uint32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const uint32_t idx = row * W + col;
if (row < H && col < W)
label[idx] = idx;
}
__global__ void
merge(uint8_t* img, int32_t* label, const uint32_t W, const uint32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const uint32_t idx = row * W + col;
if (row >= H || col >= W)
return;
uint32_t P = 0;
if (img[idx])
P |= 0x777;
if (row + 1 < H && img[idx + W])
P |= 0x777 << 4;
if (col + 1 < W && img[idx + 1])
P |= 0x777 << 1;
if (col == 0)
P &= 0xEEEE;
if (col + 1 >= W)
P &= 0x3333;
else if (col + 2 >= W)
P &= 0x7777;
if (row == 0)
P &= 0xFFF0;
if (row + 1 >= H)
P &= 0xFF;
if (P > 0) {
// If need check about top-left pixel(if flag the first bit) and hit the
// top-left pixel
if (hasBit(P, 0) && img[idx - W - 1]) {
union_(label, idx, idx - 2 * W - 2); // top left block
}
if ((hasBit(P, 1) && img[idx - W]) || (hasBit(P, 2) && img[idx - W + 1]))
union_(label, idx, idx - 2 * W); // top bottom block
if (hasBit(P, 3) && img[idx + 2 - W])
union_(label, idx, idx - 2 * W + 2); // top right block
if ((hasBit(P, 4) && img[idx - 1]) || (hasBit(P, 8) && img[idx + W - 1]))
union_(label, idx, idx - 2); // just left block
}
}
__global__ void compression(int32_t* label, const int32_t W, const int32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const uint32_t idx = row * W + col;
if (row < H && col < W)
find_n_compress(label, idx);
}
__global__ void final_labeling(
const uint8_t* img,
int32_t* label,
const int32_t W,
const int32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
const uint32_t idx = row * W + col;
if (row >= H || col >= W)
return;
int32_t y = label[idx] + 1;
if (img[idx])
label[idx] = y;
else
label[idx] = 0;
if (col + 1 < W) {
if (img[idx + 1])
label[idx + 1] = y;
else
label[idx + 1] = 0;
if (row + 1 < H) {
if (img[idx + W + 1])
label[idx + W + 1] = y;
else
label[idx + W + 1] = 0;
}
}
if (row + 1 < H) {
if (img[idx + W])
label[idx + W] = y;
else
label[idx + W] = 0;
}
}
__global__ void init_counting(
const int32_t* label,
int32_t* count_init,
const int32_t W,
const int32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
const uint32_t idx = row * W + col;
if (row >= H || col >= W)
return;
int32_t y = label[idx];
if (y > 0) {
int32_t count_idx = y - 1;
atomicAdd(count_init + count_idx, 1);
}
}
__global__ void final_counting(
const int32_t* label,
const int32_t* count_init,
int32_t* count_final,
const int32_t W,
const int32_t H) {
const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
const uint32_t idx = row * W + col;
if (row >= H || col >= W)
return;
int32_t y = label[idx];
if (y > 0) {
int32_t count_idx = y - 1;
count_final[idx] = count_init[count_idx];
} else {
count_final[idx] = 0;
}
}
} // namespace cc2d
std::vector<torch::Tensor> get_connected_componnets(
const torch::Tensor& inputs) {
AT_ASSERTM(inputs.is_cuda(), "inputs must be a CUDA tensor");
AT_ASSERTM(inputs.ndimension() == 4, "inputs must be [N, 1, H, W] shape");
AT_ASSERTM(
inputs.scalar_type() == torch::kUInt8, "inputs must be a uint8 type");
const uint32_t N = inputs.size(0);
const uint32_t C = inputs.size(1);
const uint32_t H = inputs.size(2);
const uint32_t W = inputs.size(3);
AT_ASSERTM(C == 1, "inputs must be [N, 1, H, W] shape");
AT_ASSERTM((H % 2) == 0, "height must be an even number");
AT_ASSERTM((W % 2) == 0, "width must be an even number");
// label must be uint32_t
auto label_options =
torch::TensorOptions().dtype(torch::kInt32).device(inputs.device());
torch::Tensor labels = torch::zeros({N, C, H, W}, label_options);
torch::Tensor counts_init = torch::zeros({N, C, H, W}, label_options);
torch::Tensor counts_final = torch::zeros({N, C, H, W}, label_options);
dim3 grid = dim3(
((W + 1) / 2 + BLOCK_COLS - 1) / BLOCK_COLS,
((H + 1) / 2 + BLOCK_ROWS - 1) / BLOCK_ROWS);
dim3 block = dim3(BLOCK_COLS, BLOCK_ROWS);
dim3 grid_count =
dim3((W + BLOCK_COLS) / BLOCK_COLS, (H + BLOCK_ROWS) / BLOCK_ROWS);
dim3 block_count = dim3(BLOCK_COLS, BLOCK_ROWS);
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
for (int n = 0; n < N; n++) {
uint32_t offset = n * H * W;
cc2d::init_labeling<<<grid, block, 0, stream>>>(
labels.data_ptr<int32_t>() + offset, W, H);
cc2d::merge<<<grid, block, 0, stream>>>(
inputs.data_ptr<uint8_t>() + offset,
labels.data_ptr<int32_t>() + offset,
W,
H);
cc2d::compression<<<grid, block, 0, stream>>>(
labels.data_ptr<int32_t>() + offset, W, H);
cc2d::final_labeling<<<grid, block, 0, stream>>>(
inputs.data_ptr<uint8_t>() + offset,
labels.data_ptr<int32_t>() + offset,
W,
H);
// get the counting of each pixel
cc2d::init_counting<<<grid_count, block_count, 0, stream>>>(
labels.data_ptr<int32_t>() + offset,
counts_init.data_ptr<int32_t>() + offset,
W,
H);
cc2d::final_counting<<<grid_count, block_count, 0, stream>>>(
labels.data_ptr<int32_t>() + offset,
counts_init.data_ptr<int32_t>() + offset,
counts_final.data_ptr<int32_t>() + offset,
W,
H);
}
// returned values are [labels, counts]
std::vector<torch::Tensor> outputs;
outputs.push_back(labels);
outputs.push_back(counts_final);
return outputs;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def(
"get_connected_componnets",
&get_connected_componnets,
"get_connected_componnets");
}
|