File size: 6,940 Bytes
e05eed1 98a67a0 |
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 |
// SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
#include "quad_rectify_cpu.h"
#include <iostream>
#include "../geometry.h"
#include "quad_rectify_shared.h"
using namespace std;
template<typename quads_accessor_t, typename output_accessor_t, typename scalar_t>
void quad_rectify_calc_quad_width_impl(const quads_accessor_t &quads,
output_accessor_t output,
const scalar_t outputHeight,
const scalar_t roundFactor,
const scalar_t maxWidth)
{
const int64_t numQuads = quads.size(0);
for (int64_t quadIdx = 0; quadIdx < numQuads; ++quadIdx) {
auto quadWidth = calc_quad_width(quads[quadIdx], outputHeight, roundFactor, maxWidth);
output[quadIdx] = Convert<scalar_t, int64_t>::LeftToRight(quadWidth);
}
}
template<typename quads_accessor_t, typename output_accessor_t, typename scalar_t>
void quad_rectify_cpu_forward_impl(const quads_accessor_t &quads,
output_accessor_t output,
const scalar_t imageHeight,
const scalar_t imageWidth,
bool isotropic)
{
typedef Point_<scalar_t> Point_t;
const int64_t numQuads = quads.size(0);
const int64_t outputHeight = output.size(1);
const int64_t outputWidth = output.size(2);
for (int64_t quadIdx = 0; quadIdx < numQuads; ++quadIdx) {
auto currQuad = quads[quadIdx];
scalar_t quadWidth = isotropic ? calc_quad_width<scalar_t>(currQuad, outputHeight, 1, outputWidth) : scalar_t(outputWidth);
for (int64_t row = 0; row < outputHeight; ++row) {
for (int64_t col = 0; col < outputWidth; ++col) {
Point_t outputPoint = calc_rect_value<scalar_t>(currQuad,
quadWidth,
outputHeight,
col,
row,
imageWidth,
imageHeight);
auto currOutput = output[quadIdx][row][col];
currOutput[0] = outputPoint.X;
currOutput[1] = outputPoint.Y;
}
}
}
}
/*template<typename scalar_t>
void quad_rectify_cpu_backward_impl(torch::Tensor quads,
torch::Tensor gradOutput,
torch::Tensor gradInput)
{
const int64_t batchSize = gradOutput.size(0);
const int64_t outputHeight = gradOutput.size(1);
const int64_t outputWidth = gradOutput.size(2);
auto gradInputAccess = gradInput.accessor<scalar_t, 3>();
auto gradOutputAccess = gradOutput.accessor<scalar_t, 4>();
for (int64_t batchIdx = 0; batchIdx < batchSize; ++batchIdx) {
auto batchInputAccess = gradInputAccess[batchIdx];
auto batchOutputAccess = gradOutputAccess[batchIdx];
for (int64_t rowIdx = 0; rowIdx < outputHeight; ++rowIdx) {
for (int64_t colIdx = 0; colIdx < outputWidth; ++colIdx) {
const scalar_t fRow = scalar_t(rowIdx) / outputHeight;
const scalar_t fCol = scalar_t(colIdx) / outputWidth;
const scalar_t fRowCol = fRow * fCol;
for (int64_t dim = 0; dim < 2; ++dim) {
const scalar_t dOut = batchOutputAccess[rowIdx][colIdx][dim];
const scalar_t gradIns[] = {
dOut * (fRowCol - fCol - fRow + 1),
dOut * (fCol - fRowCol),
dOut * fRowCol,
dOut * (fRow - fRowCol)
};
for (int64_t quadIdx = 0; quadIdx < 4; ++quadIdx) {
batchInputAccess[quadIdx][dim] += 2.0f * gradIns[quadIdx];
}
}
}
}
}
}*/
torch::Tensor quad_rectify_cpu_calc_quad_width(torch::Tensor quads,
int64_t outputHeight,
int64_t roundFactor,
float maxWidth)
{
auto output = torch::empty({ quads.size(0) },
quads.options().dtype(torch::kInt64));
AT_DISPATCH_FLOATING_TYPES(
quads.scalar_type(),
"quad_rectify_cpu_calc_quad_width",
([&] {
quad_rectify_calc_quad_width_impl(
quads.accessor<scalar_t, 3>(),
output.accessor<int64_t, 1>(),
Convert<scalar_t, int64_t>::RightToLeft(outputHeight),
Convert<scalar_t, int64_t>::RightToLeft(roundFactor),
Convert<scalar_t, float>::RightToLeft(maxWidth)
);
})
);
return output;
}
torch::Tensor quad_rectify_cpu_forward(torch::Tensor quads,
int64_t imageHeight,
int64_t imageWidth,
int64_t outputHeight,
int64_t outputWidth,
bool isotropic)
{
auto output = torch::empty({ quads.size(0), outputHeight, outputWidth, 2 },
quads.options());
AT_DISPATCH_FLOATING_TYPES(
quads.scalar_type(),
"quad_rectify_cpu_forward",
([&] {
quad_rectify_cpu_forward_impl(
quads.accessor<scalar_t, 3>(),
output.accessor<scalar_t, 4>(),
Convert<scalar_t, int64_t>::RightToLeft(imageHeight),
Convert<scalar_t, int64_t>::RightToLeft(imageWidth),
isotropic
);
})
);
return output;
}
torch::Tensor quad_rectify_cpu_backward(torch::Tensor quads,
torch::Tensor gradOutput,
int64_t imageHeight,
int64_t imageWidth,
bool isotropic)
{
auto gradInput = torch::zeros_like(quads);
throw std::runtime_error("Calling backward, and it's not implemented!");
/*AT_DISPATCH_FLOATING_TYPES_AND_HALF(
quads.scalar_type(),
"quad_rectify_cpu_backward",
([&] {
quad_rectify_cpu_backward_impl<scalar_t>(quads,
gradOutput,
gradInput);
})
);*/
return gradInput;
}
|