File size: 4,422 Bytes
f06f310 |
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 |
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <vector>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/Parallel.h>
#define BLOCK 16
__forceinline__ __device__ bool within_bounds(int h, int w, int H, int W) {
return h >= 0 && h < H && w >= 0 && w < W;
}
template <typename scalar_t>
__global__ void sampler_forward_kernel(
const torch::PackedTensorAccessor32<scalar_t,4,torch::RestrictPtrTraits> volume,
const torch::PackedTensorAccessor32<float,4,torch::RestrictPtrTraits> coords,
torch::PackedTensorAccessor32<scalar_t,4,torch::RestrictPtrTraits> corr,
int r)
{
// batch index
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int n = blockIdx.z;
const int h1 = volume.size(1);
const int w1 = volume.size(2);
const int w2 = volume.size(3);
if (!within_bounds(y, x, h1, w1)) {
return;
}
float x0 = coords[n][0][y][x];
float y0 = coords[n][1][y][x];
float dx = x0 - floor(x0);
float dy = y0 - floor(y0);
int rd = 2*r + 1;
for (int i=0; i<rd+1; i++) { // i is X
int x1 = static_cast<int>(floor(x0)) - r + i;
if (within_bounds(0, x1, 1, w2)) {
scalar_t s = volume[n][y][x][x1];
if (i > 0)
corr[n][i-1][y][x] += s * scalar_t(dx);
if (i < rd)
corr[n][i][y][x] += s * scalar_t((1.0f-dx));
}
}
}
template <typename scalar_t>
__global__ void sampler_backward_kernel(
const torch::PackedTensorAccessor32<float,4,torch::RestrictPtrTraits> coords,
const torch::PackedTensorAccessor32<scalar_t,4,torch::RestrictPtrTraits> corr_grad,
torch::PackedTensorAccessor32<scalar_t,4,torch::RestrictPtrTraits> volume_grad,
int r)
{
// batch index
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int n = blockIdx.z;
const int h1 = volume_grad.size(1);
const int w1 = volume_grad.size(2);
const int w2 = volume_grad.size(3);
if (!within_bounds(y, x, h1, w1)) {
return;
}
float x0 = coords[n][0][y][x];
float y0 = coords[n][1][y][x];
float dx = x0 - floor(x0);
float dy = y0 - floor(y0);
int rd = 2*r + 1;
for (int i=0; i<rd+1; i++) {
int x1 = static_cast<int>(floor(x0)) - r + i;
if (within_bounds(0, x1, 1, w2)) {
scalar_t g = 0.0;
if (i > 0)
g += corr_grad[n][i-1][y][x] * scalar_t(dx);
if (i < rd)
g += corr_grad[n][i][y][x] * scalar_t((1.0f-dx));
volume_grad[n][y][x][x1] += g;
}
}
}
std::vector<torch::Tensor> sampler_cuda_forward(
torch::Tensor volume,
torch::Tensor coords,
int radius)
{
const auto batch_size = volume.size(0);
const auto ht = volume.size(1);
const auto wd = volume.size(2);
const dim3 blocks((wd + BLOCK - 1) / BLOCK,
(ht + BLOCK - 1) / BLOCK,
batch_size);
const dim3 threads(BLOCK, BLOCK);
auto opts = volume.options();
torch::Tensor corr = torch::zeros(
{batch_size, 2*radius+1, ht, wd}, opts);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), "sampler_forward_kernel", ([&] {
sampler_forward_kernel<scalar_t><<<blocks, threads>>>(
volume.packed_accessor32<scalar_t,4,torch::RestrictPtrTraits>(),
coords.packed_accessor32<float,4,torch::RestrictPtrTraits>(),
corr.packed_accessor32<scalar_t,4,torch::RestrictPtrTraits>(),
radius);
}));
return {corr};
}
std::vector<torch::Tensor> sampler_cuda_backward(
torch::Tensor volume,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius)
{
const auto batch_size = volume.size(0);
const auto ht = volume.size(1);
const auto wd = volume.size(2);
auto volume_grad = torch::zeros_like(volume);
const dim3 blocks((wd + BLOCK - 1) / BLOCK,
(ht + BLOCK - 1) / BLOCK,
batch_size);
const dim3 threads(BLOCK, BLOCK);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), "sampler_backward_kernel", ([&] {
sampler_backward_kernel<scalar_t><<<blocks, threads>>>(
coords.packed_accessor32<float,4,torch::RestrictPtrTraits>(),
corr_grad.packed_accessor32<scalar_t,4,torch::RestrictPtrTraits>(),
volume_grad.packed_accessor32<scalar_t,4,torch::RestrictPtrTraits>(),
radius);
}));
return {volume_grad};
}
|