Files
2026-07-13 12:40:42 +08:00

2308 lines
83 KiB
C++

// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/phi/kernels/interpolate_kernel.h"
#include <array>
#include <cmath>
#include <cstdint>
#include <type_traits>
#include <vector>
#include "paddle/common/layout.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/interpolate_function.h"
namespace phi {
template <typename T>
static inline T cubic_interp(T x0, T x1, T x2, T x3, T t) {
std::array<T, 4> coeffs;
funcs::GetCubicUpsampleCoefficients<T>(coeffs.data(), t);
return x0 * coeffs[0] + x1 * coeffs[1] + x2 * coeffs[2] + x3 * coeffs[3];
}
template <typename T>
static void LinearInterpolation(const DenseTensor& input,
DenseTensor* output,
const float ratio_w,
const int in_w,
const int n,
const int c,
const int out_w,
const bool align_corners,
const int align_mode,
const DataLayout data_layout) {
auto input_t = EigenTensor<T, 3>::From(input);
auto output_t = EigenTensor<T, 3>::From(*output);
bool align_flag = (align_mode == 0 && !align_corners);
using MT = typename MPTypeTrait<T>::Type;
std::vector<int> vx_w, vx_e;
std::vector<MT> vd_w, vd_e;
vx_w.reserve(out_w);
vx_e.reserve(out_w);
vd_w.reserve(out_w);
vd_e.reserve(out_w);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int l = 0; l < out_w; l++) {
int x_w = static_cast<int>(align_flag ? (ratio_w * (l + 0.5) - 0.5)
: ratio_w * l);
x_w = (x_w > 0) ? x_w : 0; // w
int x_e = (x_w < (in_w - 1)) ? (x_w + 1) : x_w; // w_id
MT idx_src_x = ratio_w * (l + 0.5) - 0.5;
idx_src_x = (idx_src_x > 0) ? idx_src_x : 0;
MT d_w = align_flag ? idx_src_x - x_w : ratio_w * l - x_w; // w1lambda
MT d_e = 1. - d_w; // w2lambda
{
vx_w[l] = x_w;
vx_e[l] = x_e;
vd_w[l] = d_w;
vd_e[l] = d_e;
}
}
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for collapse(3)
#endif
for (int i = 0; i < n; i++) { // loop for batches
for (int j = 0; j < c; j++) { // loop for channels
for (int l = 0; l < out_w; l++) {
// linear interpolation
T out_t;
if (data_layout == DataLayout::NCHW) {
out_t =
static_cast<T>(static_cast<MT>(input_t(i, j, vx_w[l])) * vd_e[l] +
static_cast<MT>(input_t(i, j, vx_e[l])) * vd_w[l]);
output_t(i, j, l) = out_t;
} else {
out_t =
static_cast<T>(static_cast<MT>(input_t(i, vx_w[l], j)) * vd_e[l] +
static_cast<MT>(input_t(i, vx_e[l], j)) * vd_w[l]);
output_t(i, l, j) = out_t;
}
}
}
}
}
template <typename T>
static void BilinearInterpolation(const DenseTensor& input,
DenseTensor* output,
const float ratio_h,
const float ratio_w,
const int in_h,
const int in_w,
const int n,
const int c,
const int out_h,
const int out_w,
const bool align_corners,
const int align_mode,
const DataLayout data_layout) {
auto input_t = EigenTensor<T, 4>::From(input);
auto output_t = EigenTensor<T, 4>::From(*output);
bool align_flag = (align_mode == 0 && !align_corners);
using MT = typename MPTypeTrait<T>::Type;
std::vector<int> vy_n, vy_s;
std::vector<float> vd_n, vd_s;
vy_n.reserve(out_h);
vy_s.reserve(out_h);
vd_n.reserve(out_h);
vd_s.reserve(out_h);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int k = 0; k < out_h; k++) {
int y_n = static_cast<int>(align_flag ? (ratio_h * (k + 0.5) - 0.5)
: (ratio_h * static_cast<float>(k)));
y_n = (y_n > 0) ? y_n : 0;
int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1);
float idx_src_y = ratio_h * (k + 0.5) - 0.5;
idx_src_y = (idx_src_y > 0) ? idx_src_y : 0;
float d_n = align_flag
? idx_src_y - static_cast<float>(y_n)
: ratio_h * static_cast<float>(k) - static_cast<float>(y_n);
float d_s = 1.f - d_n;
{
vy_n[k] = y_n;
vy_s[k] = y_s;
vd_n[k] = d_n;
vd_s[k] = d_s;
}
}
std::vector<int> vx_w, vx_e;
std::vector<float> vd_w, vd_e;
vx_w.reserve(out_w);
vx_e.reserve(out_w);
vd_w.reserve(out_w);
vd_e.reserve(out_w);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int l = 0; l < out_w; l++) {
int x_w = (align_mode == 0 && !align_corners)
? static_cast<int>(ratio_w * (l + 0.5) - 0.5)
: static_cast<int>(ratio_w * static_cast<float>(l));
x_w = (x_w > 0) ? x_w : 0;
int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1);
float idx_src_x = ratio_w * (static_cast<float>(l) + 0.5f) - 0.5f;
idx_src_x = (idx_src_x > 0) ? idx_src_x : 0;
float d_w = align_flag
? idx_src_x - static_cast<float>(x_w)
: ratio_w * static_cast<float>(l) - static_cast<float>(x_w);
float d_e = 1.f - d_w;
{
vx_w[l] = x_w;
vx_e[l] = x_e;
vd_w[l] = d_w;
vd_e[l] = d_e;
}
}
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for collapse(4)
#endif
for (int i = 0; i < n; i++) { // loop for batches
for (int j = 0; j < c; j++) { // loop for channels
for (int k = 0; k < out_h; k++) { // loop for images
for (int l = 0; l < out_w; l++) {
// bilinear interpolation
T out_t;
if (data_layout == DataLayout::NCHW) {
out_t = static_cast<T>(
static_cast<MT>(input_t(i, j, vy_n[k], vx_w[l])) * vd_s[k] *
vd_e[l] +
static_cast<MT>(input_t(i, j, vy_s[k], vx_w[l])) * vd_n[k] *
vd_e[l] +
static_cast<MT>(input_t(i, j, vy_n[k], vx_e[l])) * vd_s[k] *
vd_w[l] +
static_cast<MT>(input_t(i, j, vy_s[k], vx_e[l])) * vd_n[k] *
vd_w[l]);
output_t(i, j, k, l) = out_t;
} else {
out_t = static_cast<T>(
static_cast<MT>(input_t(i, vy_n[k], vx_w[l], j)) * vd_s[k] *
vd_e[l] +
static_cast<MT>(input_t(i, vy_s[k], vx_w[l], j)) * vd_n[k] *
vd_e[l] +
static_cast<MT>(input_t(i, vy_n[k], vx_e[l], j)) * vd_s[k] *
vd_w[l] +
static_cast<MT>(input_t(i, vy_s[k], vx_e[l], j)) * vd_n[k] *
vd_w[l]);
output_t(i, k, l, j) = out_t;
}
}
}
}
}
}
template <typename T>
static void NearestNeighborInterpolate(const DenseTensor& input,
DenseTensor* output,
const float ratio_h,
const float ratio_w,
const int n,
const int c,
const int out_h,
const int out_w,
const bool align_corners,
const DataLayout& data_layout) {
auto input_t = EigenTensor<T, 4>::From(input);
auto output_t = EigenTensor<T, 4>::From(*output);
for (int k = 0; k < out_h; k++) { // loop for images
int in_k =
(align_corners)
? static_cast<int>(std::lround(ratio_h * static_cast<float>(k)))
: static_cast<int>(ratio_h * static_cast<float>(k));
for (int l = 0; l < out_w; l++) {
int in_l =
(align_corners)
? static_cast<int>(std::lround(ratio_w * static_cast<float>(l)))
: static_cast<int>(ratio_w * static_cast<float>(l));
for (int i = 0; i < n; i++) { // loop for batches
for (int j = 0; j < c; j++) { // loop for channels
if (data_layout == DataLayout::NCHW) {
output_t(i, j, k, l) = input_t(i, j, in_k, in_l);
} else {
output_t(i, k, l, j) = input_t(i, in_k, in_l, j);
}
}
}
}
}
}
template <typename T>
static void BicubicInterpolation(const DenseTensor& input,
DenseTensor* output,
const float ratio_h,
const float ratio_w,
const int in_h,
const int in_w,
const int n,
const int c,
const int out_h,
const int out_w,
const bool align_corners,
const DataLayout data_layout) {
auto input_t = EigenTensor<T, 4>::From(input);
auto output_t = EigenTensor<T, 4>::From(*output);
using MT = typename MPTypeTrait<T>::Type;
for (int k = 0; k < out_h; k++) { // loop for images
MT y_n = align_corners ? static_cast<MT>(ratio_h * static_cast<float>(k))
: static_cast<MT>(ratio_h * (k + 0.5) - 0.5);
int input_y = floorf(y_n);
const MT y_t = y_n - input_y;
for (int l = 0; l < out_w; l++) {
MT x_n = align_corners ? static_cast<MT>(ratio_w * static_cast<float>(l))
: static_cast<MT>(ratio_w * (l + 0.5) - 0.5);
int input_x = floorf(x_n);
const MT x_t = x_n - input_x;
for (int i = 0; i < n; i++) { // loop for batches
for (int j = 0; j < c; j++) { // loop for channels
std::array<MT, 4> coefficients;
// interp 4 times in x direction
for (int ii = 0; ii < 4; ii++) {
int access_y = std::max(std::min(input_y - 1 + ii, in_h - 1),
static_cast<int>(0));
int access_x_0 =
std::max(std::min(input_x - 1, in_w - 1), static_cast<int>(0));
int access_x_1 =
std::max(std::min(input_x + 0, in_w - 1), static_cast<int>(0));
int access_x_2 =
std::max(std::min(input_x + 1, in_w - 1), static_cast<int>(0));
int access_x_3 =
std::max(std::min(input_x + 2, in_w - 1), static_cast<int>(0));
if (data_layout == DataLayout::NCHW) {
coefficients[ii] = cubic_interp<MT>(
static_cast<MT>(input_t(i, j, access_y, access_x_0)),
static_cast<MT>(input_t(i, j, access_y, access_x_1)),
static_cast<MT>(input_t(i, j, access_y, access_x_2)),
static_cast<MT>(input_t(i, j, access_y, access_x_3)),
x_t);
} else {
coefficients[ii] = cubic_interp<MT>(
static_cast<MT>(input_t(i, access_y, access_x_0, j)),
static_cast<MT>(input_t(i, access_y, access_x_1, j)),
static_cast<MT>(input_t(i, access_y, access_x_2, j)),
static_cast<MT>(input_t(i, access_y, access_x_3, j)),
x_t);
}
}
// interp y direction
if (data_layout == DataLayout::NCHW) {
output_t(i, j, k, l) =
static_cast<T>(cubic_interp<MT>(coefficients[0],
coefficients[1],
coefficients[2],
coefficients[3],
y_t));
} else {
output_t(i, k, l, j) =
static_cast<T>(cubic_interp<MT>(coefficients[0],
coefficients[1],
coefficients[2],
coefficients[3],
y_t));
}
}
}
}
}
}
template <typename T>
static void TrilinearInterpolation(const DenseTensor& input,
DenseTensor* output,
const float ratio_d,
const float ratio_h,
const float ratio_w,
const int in_d,
const int in_h,
const int in_w,
const int n,
const int c,
const int out_d,
const int out_h,
const int out_w,
const bool align_corners,
const int align_mode,
const DataLayout& data_layout) {
auto input_t = EigenTensor<T, 5>::From(input);
auto output_t = EigenTensor<T, 5>::From(*output);
bool align_flag = (align_mode == 0 && !align_corners);
using MT = typename MPTypeTrait<T>::Type;
std::vector<int> vt_f, vt_b;
std::vector<float> vd_f, vd_b;
vt_f.reserve(out_d);
vt_b.reserve(out_d);
vd_f.reserve(out_d);
vd_b.reserve(out_d);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int j = 0; j < out_d; j++) {
int t_f = align_flag ? static_cast<int>(ratio_d * (j + 0.5) - 0.5)
: static_cast<int>(ratio_d * static_cast<float>(j));
t_f = (t_f > 0) ? t_f : 0;
int t_b = (t_f + 1) < (in_d - 1) ? (t_f + 1) : (in_d - 1);
float idx_src_t = ratio_d * (static_cast<float>(j) + 0.5f) - 0.5f;
idx_src_t = (idx_src_t > 0) ? idx_src_t : 0;
float d_f = align_flag
? idx_src_t - static_cast<float>(t_f)
: ratio_d * static_cast<float>(j) - static_cast<float>(t_f);
float d_b = 1.f - d_f;
{
vt_f[j] = t_f;
vt_b[j] = t_b;
vd_f[j] = d_f;
vd_b[j] = d_b;
}
}
std::vector<int> vy_n, vy_s;
std::vector<float> vd_n, vd_s;
vy_n.reserve(out_h);
vy_s.reserve(out_h);
vd_n.reserve(out_h);
vd_s.reserve(out_h);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int k = 0; k < out_h; k++) {
int y_n = align_flag ? static_cast<int>(ratio_h * (k + 0.5) - 0.5)
: static_cast<int>(ratio_h * static_cast<float>(k));
y_n = (y_n > 0) ? y_n : 0;
int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1);
float idx_src_y = ratio_h * (static_cast<float>(k) + 0.5f) - 0.5f;
idx_src_y = (idx_src_y > 0) ? idx_src_y : 0;
float d_n = align_flag
? idx_src_y - static_cast<float>(y_n)
: ratio_h * static_cast<float>(k) - static_cast<float>(y_n);
float d_s = 1.f - d_n;
{
vy_n[k] = y_n;
vy_s[k] = y_s;
vd_n[k] = d_n;
vd_s[k] = d_s;
}
}
std::vector<int> vx_w, vx_e;
std::vector<float> vd_w, vd_e;
vx_w.reserve(out_w);
vx_e.reserve(out_w);
vd_w.reserve(out_w);
vd_e.reserve(out_w);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int l = 0; l < out_w; l++) {
int x_w = (align_mode == 0 && !align_corners)
? static_cast<int>(ratio_w * (l + 0.5) - 0.5)
: static_cast<int>(ratio_w * static_cast<float>(l));
x_w = (x_w > 0) ? x_w : 0;
int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1);
float idx_src_x = ratio_w * (static_cast<float>(l) + 0.5f) - 0.5f;
idx_src_x = (idx_src_x > 0) ? idx_src_x : 0;
float d_w = align_flag
? idx_src_x - static_cast<float>(x_w)
: ratio_w * static_cast<float>(l) - static_cast<float>(x_w);
float d_e = 1.f - d_w;
{
vx_w[l] = x_w;
vx_e[l] = x_e;
vd_w[l] = d_w;
vd_e[l] = d_e;
}
}
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for collapse(5)
#endif
for (int b = 0; b < n; b++) { // loop for batches
for (int i = 0; i < c; i++) { // loop for channels
for (int j = 0; j < out_d; j++) { // loop for D, H, W
for (int k = 0; k < out_h; k++) {
for (int l = 0; l < out_w; l++) {
// trilinear interpolation
if (data_layout == DataLayout::NCHW) {
T out_t = static_cast<T>(
static_cast<MT>(input_t(b, i, vt_f[j], vy_n[k], vx_w[l])) *
vd_b[j] * vd_s[k] * vd_e[l] +
static_cast<MT>(input_t(b, i, vt_f[j], vy_n[k], vx_e[l])) *
vd_b[j] * vd_s[k] * vd_w[l] +
static_cast<MT>(input_t(b, i, vt_f[j], vy_s[k], vx_w[l])) *
vd_b[j] * vd_n[k] * vd_e[l] +
static_cast<MT>(input_t(b, i, vt_f[j], vy_s[k], vx_e[l])) *
vd_b[j] * vd_n[k] * vd_w[l] +
static_cast<MT>(input_t(b, i, vt_b[j], vy_n[k], vx_w[l])) *
vd_f[j] * vd_s[k] * vd_e[l] +
static_cast<MT>(input_t(b, i, vt_b[j], vy_n[k], vx_e[l])) *
vd_f[j] * vd_s[k] * vd_w[l] +
static_cast<MT>(input_t(b, i, vt_b[j], vy_s[k], vx_w[l])) *
vd_f[j] * vd_n[k] * vd_e[l] +
static_cast<MT>(input_t(b, i, vt_b[j], vy_s[k], vx_e[l])) *
vd_f[j] * vd_n[k] * vd_w[l]);
output_t(b, i, j, k, l) = out_t;
} else {
T out_t = static_cast<T>(
static_cast<MT>(input_t(b, vt_f[j], vy_n[k], vx_w[l], i)) *
vd_b[j] * vd_s[k] * vd_e[l] +
static_cast<MT>(input_t(b, vt_f[j], vy_n[k], vx_e[l], i)) *
vd_b[j] * vd_s[k] * vd_w[l] +
static_cast<MT>(input_t(b, vt_f[j], vy_s[k], vx_w[l], i)) *
vd_b[j] * vd_n[k] * vd_e[l] +
static_cast<MT>(input_t(b, vt_f[j], vy_s[k], vx_e[l], i)) *
vd_b[j] * vd_n[k] * vd_w[l] +
static_cast<MT>(input_t(b, vt_b[j], vy_n[k], vx_w[l], i)) *
vd_f[j] * vd_s[k] * vd_e[l] +
static_cast<MT>(input_t(b, vt_b[j], vy_n[k], vx_e[l], i)) *
vd_f[j] * vd_s[k] * vd_w[l] +
static_cast<MT>(input_t(b, vt_b[j], vy_s[k], vx_w[l], i)) *
vd_f[j] * vd_n[k] * vd_e[l] +
static_cast<MT>(input_t(b, vt_b[j], vy_s[k], vx_e[l], i)) *
vd_f[j] * vd_n[k] * vd_w[l]);
output_t(b, j, k, l, i) = out_t;
}
}
}
}
}
}
}
template <typename T>
static void NearestNeighbor3DInterpolate(const DenseTensor& input,
DenseTensor* output,
const float ratio_d,
const float ratio_h,
const float ratio_w,
const int n,
const int c,
const int out_d,
const int out_h,
const int out_w,
const bool align_corners,
const DataLayout& data_layout) {
auto input_t = EigenTensor<T, 5>::From(input);
auto output_t = EigenTensor<T, 5>::From(*output);
for (int d = 0; d < out_d; d++) { // loop for images
int in_d =
(align_corners)
? static_cast<int>(std::lround(ratio_d * static_cast<float>(d)))
: static_cast<int>(ratio_d * static_cast<float>(d));
for (int k = 0; k < out_h; k++) {
int in_k =
(align_corners)
? static_cast<int>(std::lround(ratio_h * static_cast<float>(k)))
: static_cast<int>(ratio_h * static_cast<float>(k));
for (int l = 0; l < out_w; l++) {
int in_l =
(align_corners)
? static_cast<int>(std::lround(ratio_w * static_cast<float>(l)))
: static_cast<int>(ratio_w * static_cast<float>(l));
for (int i = 0; i < n; i++) { // loop for batches
for (int j = 0; j < c; j++) { // loop for channels
if (data_layout == DataLayout::NCHW) {
output_t(i, j, d, k, l) = input_t(i, j, in_d, in_k, in_l);
} else { // NDHWC
output_t(i, d, k, l, j) = input_t(i, in_d, in_k, in_l, j);
}
}
}
}
}
}
}
template <typename T, typename Context>
static void Interpolate1DCPUFwd(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout_str,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const DataLayout data_layout = StringToDataLayout(data_layout_str);
int64_t n = 0, c = 0, in_d = 0, in_h = 0, in_w = 0;
funcs::ExtractNCDWH(x.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
double scale_w = -1.;
if (size_tensor && !size_tensor->empty()) {
// have size tensor
auto new_size = funcs::get_new_shape(size_tensor.get());
out_w = new_size[0];
} else {
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
scale_w = scale_data[0];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
} else {
if (!scale.empty()) {
scale_w = scale[0];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
}
}
if (scale_w > 0.) {
out_w = static_cast<int>(in_w * scale_w); // NOLINT
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_w = out_size_data[0];
}
}
PADDLE_ENFORCE_GT(
out_w,
0,
errors::InvalidArgument("out_w in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
DDim dim_out;
if (data_layout == DataLayout::NCHW) {
dim_out = {n, c, out_w};
} else {
dim_out = {n, out_w, c};
}
output->Resize(dim_out);
dev_ctx.template Alloc<T>(output);
if (in_w == out_w) {
Copy(dev_ctx, x, dev_ctx.GetPlace(), false, output);
return;
}
float ratio_w =
funcs::AreaPixelComputeScale<double>(in_w, out_w, align_corners, scale_w);
if ("linear" == interp_method) {
LinearInterpolation<T>(x,
output,
ratio_w,
in_w,
n,
c,
out_w,
align_corners,
align_mode,
data_layout);
}
}
template <typename T, typename Context>
static void Interpolate2DCPUFwd(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout_str,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const DataLayout data_layout = StringToDataLayout(data_layout_str);
int64_t n = 0, c = 0, in_d = 0, in_h = 0, in_w = 0;
funcs::ExtractNCDWH(x.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
double scale_h = -1;
double scale_w = -1;
if (size_tensor && !size_tensor->empty()) {
// have size tensor
auto new_size = funcs::get_new_shape(size_tensor.get());
out_h = new_size[0];
out_w = new_size[1];
} else {
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
} else {
if (scale.size() > 1) {
scale_h = scale[0];
scale_w = scale[1];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
}
}
if (scale_h > 0. && scale_w > 0.) {
out_h = static_cast<int>(in_h * scale_h); // NOLINT
out_w = static_cast<int>(in_w * scale_w); // NOLINT
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
PADDLE_ENFORCE_GT(
out_h,
0,
errors::InvalidArgument("out_h in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
errors::InvalidArgument("out_w in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
DDim dim_out;
if (data_layout == DataLayout::NCHW) {
dim_out = {n, c, out_h, out_w};
} else {
dim_out = {n, out_h, out_w, c};
}
output->Resize(dim_out);
dev_ctx.template Alloc<T>(output);
if (in_h == out_h && in_w == out_w) {
Copy(dev_ctx, x, dev_ctx.GetPlace(), false, output);
return;
}
float ratio_h =
funcs::AreaPixelComputeScale<float>(in_h, out_h, align_corners, scale_h);
float ratio_w =
funcs::AreaPixelComputeScale<float>(in_w, out_w, align_corners, scale_w);
// TODO(zrr1999): to align xpu
if (out_h <= 1) {
ratio_h = 0;
}
if (out_w <= 1) {
ratio_w = 0;
}
if ("bilinear" == interp_method) {
BilinearInterpolation<T>(x,
output,
ratio_h,
ratio_w,
in_h,
in_w,
n,
c,
out_h,
out_w,
align_corners,
align_mode,
data_layout);
} else if ("nearest" == interp_method) {
NearestNeighborInterpolate<T>(x,
output,
ratio_h,
ratio_w,
n,
c,
out_h,
out_w,
align_corners,
data_layout);
} else if ("bicubic" == interp_method) {
BicubicInterpolation<T>(x,
output,
ratio_h,
ratio_w,
in_h,
in_w,
n,
c,
out_h,
out_w,
align_corners,
data_layout);
}
}
template <typename T, typename Context>
static void Interpolate3DCPUFwd(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout_str,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const DataLayout data_layout = StringToDataLayout(data_layout_str);
int64_t n = 0, c = 0, in_d = 0, in_h = 0, in_w = 0;
funcs::ExtractNCDWH(x.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
double scale_d = -1;
double scale_h = -1;
double scale_w = -1;
if (size_tensor && !size_tensor->empty()) {
// have size tensor
auto new_size = funcs::get_new_shape(size_tensor.get());
out_d = new_size[0];
out_h = new_size[1];
out_w = new_size[2];
} else {
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
if (scale_data.size() > 1) {
scale_d = scale_data[0];
scale_h = scale_data[1];
scale_w = scale_data[2];
} else {
scale_d = scale_data[0];
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
PADDLE_ENFORCE_EQ(
scale_d > 0,
true,
errors::InvalidArgument(
"The scale_d in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_d));
} else {
if (scale.size() > 1) {
scale_d = scale[0];
scale_h = scale[1];
scale_w = scale[2];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
PADDLE_ENFORCE_EQ(
scale_d > 0,
true,
errors::InvalidArgument(
"The scale_d in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_d));
}
}
if (scale_w > 0. && scale_h > 0. && scale_d > 0.) {
out_d = static_cast<int>(in_d * scale_d); // NOLINT
out_h = static_cast<int>(in_h * scale_h); // NOLINT
out_w = static_cast<int>(in_w * scale_w); // NOLINT
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_d = out_size_data[0];
out_h = out_size_data[1];
out_w = out_size_data[2];
}
}
PADDLE_ENFORCE_GT(
out_d,
0,
errors::InvalidArgument("out_d in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_h,
0,
errors::InvalidArgument("out_h in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
errors::InvalidArgument("out_w in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
DDim dim_out;
if (data_layout == DataLayout::NCHW) {
dim_out = {n, c, out_d, out_h, out_w};
} else {
dim_out = {n, out_d, out_h, out_w, c};
}
output->Resize(dim_out);
dev_ctx.template Alloc<T>(output);
if (in_d == out_d && in_h == out_h && in_w == out_w) {
Copy(dev_ctx, x, dev_ctx.GetPlace(), false, output);
return;
}
float ratio_d =
funcs::AreaPixelComputeScale<float>(in_d, out_d, align_corners, scale_d);
float ratio_h =
funcs::AreaPixelComputeScale<float>(in_h, out_h, align_corners, scale_h);
float ratio_w =
funcs::AreaPixelComputeScale<float>(in_w, out_w, align_corners, scale_w);
if ("trilinear" == interp_method) {
TrilinearInterpolation<T>(x,
output,
ratio_d,
ratio_h,
ratio_w,
in_d,
in_h,
in_w,
n,
c,
out_d,
out_h,
out_w,
align_corners,
align_mode,
data_layout);
} else if ("nearest" == interp_method) {
NearestNeighbor3DInterpolate<T>(x,
output,
ratio_d,
ratio_h,
ratio_w,
n,
c,
out_d,
out_h,
out_w,
align_corners,
data_layout);
}
}
template <typename T, typename Context>
void InterpolateKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
if (x.numel() == 0) {
dev_ctx.template Alloc<T>(output);
return;
}
auto input_dims = x.dims();
if (input_dims.size() == 3) { // 1D interpolation
Interpolate1DCPUFwd<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
} else if (input_dims.size() == 4) { // 2D interpolation
Interpolate2DCPUFwd<T>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
} else if (input_dims.size() == 5) { // 3D interpolation
Interpolate3DCPUFwd<T>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
}
template <typename T, typename Context>
void BilinearInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void LegacyBilinearInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
float scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const auto& dim_x = x.dims();
std::vector<double> scale_vec;
if (scale > 0) {
for (int i = 0; i < dim_x.size() - 2; i++) {
scale_vec.push_back(scale);
}
}
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale_vec,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void NearestInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void LegacyNearestInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
float scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const auto& dim_x = x.dims();
std::vector<double> scale_vec;
if (scale > 0) {
for (int i = 0; i < dim_x.size() - 2; i++) {
scale_vec.push_back(scale);
}
}
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale_vec,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void TrilinearInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void LinearInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void BicubicInterpKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
// =====================================================================
// CPU Antialias Interpolation Forward Implementation
// Separable 2-pass AA interpolation matching PyTorch's behavior exactly.
// =====================================================================
// CPU weight computation for antialias interpolation.
// Matches the GPU ComputeWeights function and PyTorch's weight computation.
template <typename WT, typename InterpFilter>
static void ComputeAAWeightsCPU(WT* wt_ptr,
const WT scale,
int interp_size,
const InterpFilter& interp_filter,
WT xmin_m_center,
int xsize) {
WT invscale = (scale >= static_cast<WT>(1.0)) ? static_cast<WT>(1.0) / scale
: static_cast<WT>(1.0);
WT total_w = static_cast<WT>(0.0);
int j = 0;
for (j = 0; j < xsize; j++) {
WT w = interp_filter((j + xmin_m_center + static_cast<WT>(0.5)) * invscale);
wt_ptr[j] = w;
total_w += w;
}
for (j = 0; j < xsize; j++) {
if (total_w != static_cast<WT>(0.0)) {
wt_ptr[j] /= total_w;
}
}
for (; j < interp_size; j++) {
wt_ptr[j] = static_cast<WT>(0.0);
}
}
// CPU weight span computation matching the GPU ComputeWeightsSpan.
template <typename WT>
static void ComputeAAWeightsSpanCPU(const int i,
const int input_size,
const WT scale,
const WT support,
int* xmin,
int* xsize,
WT* center) {
*center = scale * (i + static_cast<WT>(0.5));
*xmin = std::max(
static_cast<int>(std::floor(*center - support + static_cast<WT>(0.5))),
0);
*xsize = std::min(static_cast<int>(
std::floor(*center + support + static_cast<WT>(0.5))),
input_size) -
*xmin;
}
// Single dimension AA interpolation for float types on CPU.
// Computes weighted sum: sum(src[j] * weights[j]) for j in [0, size).
template <typename T, typename WT>
static WT InterpolateAASingleDimCPU(const T* src, const WT* weights, int size) {
WT output = static_cast<WT>(src[0]) * weights[0];
for (int j = 1; j < size; j++) {
output += static_cast<WT>(src[j]) * weights[j];
}
return output;
}
// Forward pass: separable 2-pass AA interpolation for float types, NCHW.
// Pass 1 (horizontal): input [N,C,H_in,W_in] -> temp [N,C,H_in,W_out]
// Pass 2 (vertical): temp [N,C,H_in,W_out] -> output [N,C,H_out,W_out]
template <typename T, typename InterpFilter>
static void AAInterpolation2DCPU_NCHW(const T* input_data,
T* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const InterpFilter& filter) {
// Use MPTypeTrait to match GPU: float for float/float16/bfloat16, double for
// double
using WT = typename MPTypeTrait<T>::Type;
WT scale_h = static_cast<WT>(ratio_h);
WT scale_w = static_cast<WT>(ratio_w);
const WT half = static_cast<WT>(0.5);
const WT support_h = (scale_h >= static_cast<WT>(1.0))
? (filter.size * half) * scale_h
: filter.size * half;
const WT support_w = (scale_w >= static_cast<WT>(1.0))
? (filter.size * half) * scale_w
: filter.size * half;
const int interp_height = static_cast<int>(std::ceil(support_h)) * 2 + 1;
const int interp_width = static_cast<int>(std::ceil(support_w)) * 2 + 1;
// Allocate temporary buffer for intermediate result [N, C, H_in, W_out]
// and weight arrays
std::vector<T> temp(static_cast<size_t>(n) * c * in_h * out_w);
std::vector<WT> wx(interp_width);
std::vector<WT> wy(interp_height);
// Pre-compute horizontal weights and spans for each output column
struct SpanInfo {
int xmin;
int xsize;
WT center;
};
std::vector<SpanInfo> h_spans(out_w);
std::vector<std::vector<WT>> h_weights(out_w);
for (int ow = 0; ow < out_w; ow++) {
ComputeAAWeightsSpanCPU<WT>(ow,
in_w,
scale_w,
support_w,
&h_spans[ow].xmin,
&h_spans[ow].xsize,
&h_spans[ow].center);
h_weights[ow].resize(interp_width);
ComputeAAWeightsCPU<WT>(
h_weights[ow].data(),
scale_w,
interp_width,
filter,
static_cast<WT>(h_spans[ow].xmin) - h_spans[ow].center,
h_spans[ow].xsize);
}
// Pre-compute vertical weights and spans for each output row
std::vector<SpanInfo> v_spans(out_h);
std::vector<std::vector<WT>> v_weights(out_h);
for (int oh = 0; oh < out_h; oh++) {
ComputeAAWeightsSpanCPU<WT>(oh,
in_h,
scale_h,
support_h,
&v_spans[oh].xmin,
&v_spans[oh].xsize,
&v_spans[oh].center);
v_weights[oh].resize(interp_height);
ComputeAAWeightsCPU<WT>(
v_weights[oh].data(),
scale_h,
interp_height,
filter,
static_cast<WT>(v_spans[oh].xmin) - v_spans[oh].center,
v_spans[oh].xsize);
}
// Pass 1: Horizontal interpolation
// For each (batch, channel, input_row), interpolate across width
for (int64_t nc_idx = 0; nc_idx < n * c; nc_idx++) {
for (int ih = 0; ih < in_h; ih++) {
const T* in_row = input_data + nc_idx * in_h * in_w + ih * in_w;
T* temp_row = temp.data() + nc_idx * in_h * out_w + ih * out_w;
for (int ow = 0; ow < out_w; ow++) {
int xmin = h_spans[ow].xmin;
int xsize = h_spans[ow].xsize;
const WT* wts = h_weights[ow].data();
WT result = static_cast<WT>(0);
for (int j = 0; j < xsize; j++) {
result += static_cast<WT>(in_row[xmin + j]) * wts[j];
}
temp_row[ow] = static_cast<T>(result);
}
}
}
// Pass 2: Vertical interpolation
// For each (batch, channel, output_col), interpolate across height
for (int64_t nc_idx = 0; nc_idx < n * c; nc_idx++) {
for (int oh = 0; oh < out_h; oh++) {
int ymin = v_spans[oh].xmin;
int ysize = v_spans[oh].xsize;
const WT* wts = v_weights[oh].data();
T* out_row = output_data + nc_idx * out_h * out_w + oh * out_w;
for (int ow = 0; ow < out_w; ow++) {
WT result = static_cast<WT>(0);
for (int j = 0; j < ysize; j++) {
const T* temp_row =
temp.data() + nc_idx * in_h * out_w + (ymin + j) * out_w;
result += static_cast<WT>(temp_row[ow]) * wts[j];
}
out_row[ow] = static_cast<T>(result);
}
}
}
}
// Forward pass: separable 2-pass AA interpolation for float types, NHWC.
template <typename T, typename InterpFilter>
static void AAInterpolation2DCPU_NHWC(const T* input_data,
T* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const InterpFilter& filter) {
// Use MPTypeTrait to match GPU: float for float/float16/bfloat16, double for
// double
using WT = typename MPTypeTrait<T>::Type;
WT scale_h = static_cast<WT>(ratio_h);
WT scale_w = static_cast<WT>(ratio_w);
const WT half = static_cast<WT>(0.5);
const WT support_h = (scale_h >= static_cast<WT>(1.0))
? (filter.size * half) * scale_h
: filter.size * half;
const WT support_w = (scale_w >= static_cast<WT>(1.0))
? (filter.size * half) * scale_w
: filter.size * half;
const int interp_height = static_cast<int>(std::ceil(support_h)) * 2 + 1;
const int interp_width = static_cast<int>(std::ceil(support_w)) * 2 + 1;
// Temporary buffer: [N, H_in, W_out, C]
std::vector<T> temp(static_cast<size_t>(n) * in_h * out_w * c);
// Pre-compute horizontal weights
struct SpanInfo {
int xmin;
int xsize;
WT center;
};
std::vector<SpanInfo> h_spans(out_w);
std::vector<std::vector<WT>> h_weights(out_w);
for (int ow = 0; ow < out_w; ow++) {
ComputeAAWeightsSpanCPU<WT>(ow,
in_w,
scale_w,
support_w,
&h_spans[ow].xmin,
&h_spans[ow].xsize,
&h_spans[ow].center);
h_weights[ow].resize(interp_width);
ComputeAAWeightsCPU<WT>(
h_weights[ow].data(),
scale_w,
interp_width,
filter,
static_cast<WT>(h_spans[ow].xmin) - h_spans[ow].center,
h_spans[ow].xsize);
}
// Pre-compute vertical weights
std::vector<SpanInfo> v_spans(out_h);
std::vector<std::vector<WT>> v_weights(out_h);
for (int oh = 0; oh < out_h; oh++) {
ComputeAAWeightsSpanCPU<WT>(oh,
in_h,
scale_h,
support_h,
&v_spans[oh].xmin,
&v_spans[oh].xsize,
&v_spans[oh].center);
v_weights[oh].resize(interp_height);
ComputeAAWeightsCPU<WT>(
v_weights[oh].data(),
scale_h,
interp_height,
filter,
static_cast<WT>(v_spans[oh].xmin) - v_spans[oh].center,
v_spans[oh].xsize);
}
// Pass 1: Horizontal - input [N,H_in,W_in,C] -> temp [N,H_in,W_out,C]
for (int64_t bi = 0; bi < n; bi++) {
for (int ih = 0; ih < in_h; ih++) {
for (int ow = 0; ow < out_w; ow++) {
int xmin = h_spans[ow].xmin;
int xsize = h_spans[ow].xsize;
const WT* wts = h_weights[ow].data();
for (int64_t ch = 0; ch < c; ch++) {
WT result = static_cast<WT>(0);
for (int j = 0; j < xsize; j++) {
int64_t in_idx = ((bi * in_h + ih) * in_w + (xmin + j)) * c + ch;
result += static_cast<WT>(input_data[in_idx]) * wts[j];
}
int64_t temp_idx = ((bi * in_h + ih) * out_w + ow) * c + ch;
temp[temp_idx] = static_cast<T>(result);
}
}
}
}
// Pass 2: Vertical - temp [N,H_in,W_out,C] -> output [N,H_out,W_out,C]
for (int64_t bi = 0; bi < n; bi++) {
for (int oh = 0; oh < out_h; oh++) {
int ymin = v_spans[oh].xmin;
int ysize = v_spans[oh].xsize;
const WT* wts = v_weights[oh].data();
for (int ow = 0; ow < out_w; ow++) {
for (int64_t ch = 0; ch < c; ch++) {
WT result = static_cast<WT>(0);
for (int j = 0; j < ysize; j++) {
int64_t temp_idx = ((bi * in_h + (ymin + j)) * out_w + ow) * c + ch;
result += static_cast<WT>(temp[temp_idx]) * wts[j];
}
int64_t out_idx = ((bi * out_h + oh) * out_w + ow) * c + ch;
output_data[out_idx] = static_cast<T>(result);
}
}
}
}
}
// Specialization for uint8_t: uses double weights, int16 quantization,
// int32 accumulation -- matching PyTorch's Pillow-compatible uint8 path.
template <typename InterpFilter>
static void AAInterpolation2DCPU_NCHW_UInt8(const uint8_t* input_data,
uint8_t* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const InterpFilter& filter) {
using WT = double;
WT scale_h = static_cast<WT>(ratio_h);
WT scale_w = static_cast<WT>(ratio_w);
const WT half = 0.5;
const WT support_h =
(scale_h >= 1.0) ? (filter.size * half) * scale_h : filter.size * half;
const WT support_w =
(scale_w >= 1.0) ? (filter.size * half) * scale_w : filter.size * half;
const int interp_height = static_cast<int>(std::ceil(support_h)) * 2 + 1;
const int interp_width = static_cast<int>(std::ceil(support_w)) * 2 + 1;
struct SpanInfo {
int xmin;
int xsize;
WT center;
};
// Helper: compute double weights, then quantize to int16 as PyTorch does
auto compute_int16_weights = [&](const std::vector<WT>& dbl_weights,
int xsize,
std::vector<int16_t>& i16_weights,
unsigned int& precision) {
// Find maximum weight
WT wt_max = 0.0;
for (int j = 0; j < xsize; j++) {
WT aw = dbl_weights[j] < 0 ? -dbl_weights[j] : dbl_weights[j];
if (aw > wt_max) wt_max = aw;
}
// Find max precision P such that round(max_weight * 2^(P+1)) < 2^15
unsigned int P = 0;
for (P = 0; P < 22; ++P) {
int next_value = static_cast<int>(0.5 + wt_max * (1 << (P + 1)));
if (next_value >= (1 << 15)) break;
}
precision = P;
i16_weights.resize(xsize);
for (int j = 0; j < xsize; j++) {
i16_weights[j] =
static_cast<int16_t>(std::round(dbl_weights[j] * (1 << P)));
}
};
// Pre-compute horizontal weights (double) and quantized int16 weights
std::vector<SpanInfo> h_spans(out_w);
std::vector<std::vector<WT>> h_dbl_weights(out_w);
std::vector<std::vector<int16_t>> h_i16_weights(out_w);
std::vector<unsigned int> h_precision(out_w);
for (int ow = 0; ow < out_w; ow++) {
ComputeAAWeightsSpanCPU<WT>(ow,
in_w,
scale_w,
support_w,
&h_spans[ow].xmin,
&h_spans[ow].xsize,
&h_spans[ow].center);
h_dbl_weights[ow].resize(interp_width);
ComputeAAWeightsCPU<WT>(
h_dbl_weights[ow].data(),
scale_w,
interp_width,
filter,
static_cast<WT>(h_spans[ow].xmin) - h_spans[ow].center,
h_spans[ow].xsize);
compute_int16_weights(h_dbl_weights[ow],
h_spans[ow].xsize,
h_i16_weights[ow],
h_precision[ow]);
}
// Pre-compute vertical weights
std::vector<SpanInfo> v_spans(out_h);
std::vector<std::vector<WT>> v_dbl_weights(out_h);
std::vector<std::vector<int16_t>> v_i16_weights(out_h);
std::vector<unsigned int> v_precision(out_h);
for (int oh = 0; oh < out_h; oh++) {
ComputeAAWeightsSpanCPU<WT>(oh,
in_h,
scale_h,
support_h,
&v_spans[oh].xmin,
&v_spans[oh].xsize,
&v_spans[oh].center);
v_dbl_weights[oh].resize(interp_height);
ComputeAAWeightsCPU<WT>(
v_dbl_weights[oh].data(),
scale_h,
interp_height,
filter,
static_cast<WT>(v_spans[oh].xmin) - v_spans[oh].center,
v_spans[oh].xsize);
compute_int16_weights(v_dbl_weights[oh],
v_spans[oh].xsize,
v_i16_weights[oh],
v_precision[oh]);
}
// Temporary buffer [N, C, H_in, W_out] as uint8
std::vector<uint8_t> temp(static_cast<size_t>(n) * c * in_h * out_w);
// Pass 1: Horizontal interpolation with int16 weights / int32 accumulation
for (int64_t nc_idx = 0; nc_idx < n * c; nc_idx++) {
for (int ih = 0; ih < in_h; ih++) {
const uint8_t* in_row = input_data + nc_idx * in_h * in_w + ih * in_w;
uint8_t* temp_row = temp.data() + nc_idx * in_h * out_w + ih * out_w;
for (int ow = 0; ow < out_w; ow++) {
int xmin = h_spans[ow].xmin;
int xsize = h_spans[ow].xsize;
unsigned int P = h_precision[ow];
const int16_t* i16w = h_i16_weights[ow].data();
int32_t accum = 1 << (P > 0 ? P - 1 : 0); // rounding bias
for (int j = 0; j < xsize; j++) {
accum += static_cast<int32_t>(in_row[xmin + j]) *
static_cast<int32_t>(i16w[j]);
}
int32_t result = accum >> P;
temp_row[ow] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
}
}
}
// Pass 2: Vertical interpolation
for (int64_t nc_idx = 0; nc_idx < n * c; nc_idx++) {
for (int oh = 0; oh < out_h; oh++) {
int ymin = v_spans[oh].xmin;
int ysize = v_spans[oh].xsize;
unsigned int P = v_precision[oh];
const int16_t* i16w = v_i16_weights[oh].data();
uint8_t* out_row = output_data + nc_idx * out_h * out_w + oh * out_w;
for (int ow = 0; ow < out_w; ow++) {
int32_t accum = 1 << (P > 0 ? P - 1 : 0);
for (int j = 0; j < ysize; j++) {
const uint8_t* temp_row =
temp.data() + nc_idx * in_h * out_w + (ymin + j) * out_w;
accum += static_cast<int32_t>(temp_row[ow]) *
static_cast<int32_t>(i16w[j]);
}
int32_t result = accum >> P;
out_row[ow] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
}
}
}
}
// NHWC variant for uint8
template <typename InterpFilter>
static void AAInterpolation2DCPU_NHWC_UInt8(const uint8_t* input_data,
uint8_t* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const InterpFilter& filter) {
using WT = double;
WT scale_h = static_cast<WT>(ratio_h);
WT scale_w = static_cast<WT>(ratio_w);
const WT half = 0.5;
const WT support_h =
(scale_h >= 1.0) ? (filter.size * half) * scale_h : filter.size * half;
const WT support_w =
(scale_w >= 1.0) ? (filter.size * half) * scale_w : filter.size * half;
const int interp_height = static_cast<int>(std::ceil(support_h)) * 2 + 1;
const int interp_width = static_cast<int>(std::ceil(support_w)) * 2 + 1;
struct SpanInfo {
int xmin;
int xsize;
WT center;
};
auto compute_int16_weights = [&](const std::vector<WT>& dbl_weights,
int xsize,
std::vector<int16_t>& i16_weights,
unsigned int& precision) {
WT wt_max = 0.0;
for (int j = 0; j < xsize; j++) {
WT aw = dbl_weights[j] < 0 ? -dbl_weights[j] : dbl_weights[j];
if (aw > wt_max) wt_max = aw;
}
unsigned int P = 0;
for (P = 0; P < 22; ++P) {
int next_value = static_cast<int>(0.5 + wt_max * (1 << (P + 1)));
if (next_value >= (1 << 15)) break;
}
precision = P;
i16_weights.resize(xsize);
for (int j = 0; j < xsize; j++) {
i16_weights[j] =
static_cast<int16_t>(std::round(dbl_weights[j] * (1 << P)));
}
};
std::vector<SpanInfo> h_spans(out_w);
std::vector<std::vector<WT>> h_dbl_weights(out_w);
std::vector<std::vector<int16_t>> h_i16_weights(out_w);
std::vector<unsigned int> h_precision(out_w);
for (int ow = 0; ow < out_w; ow++) {
ComputeAAWeightsSpanCPU<WT>(ow,
in_w,
scale_w,
support_w,
&h_spans[ow].xmin,
&h_spans[ow].xsize,
&h_spans[ow].center);
h_dbl_weights[ow].resize(interp_width);
ComputeAAWeightsCPU<WT>(
h_dbl_weights[ow].data(),
scale_w,
interp_width,
filter,
static_cast<WT>(h_spans[ow].xmin) - h_spans[ow].center,
h_spans[ow].xsize);
compute_int16_weights(h_dbl_weights[ow],
h_spans[ow].xsize,
h_i16_weights[ow],
h_precision[ow]);
}
std::vector<SpanInfo> v_spans(out_h);
std::vector<std::vector<WT>> v_dbl_weights(out_h);
std::vector<std::vector<int16_t>> v_i16_weights(out_h);
std::vector<unsigned int> v_precision(out_h);
for (int oh = 0; oh < out_h; oh++) {
ComputeAAWeightsSpanCPU<WT>(oh,
in_h,
scale_h,
support_h,
&v_spans[oh].xmin,
&v_spans[oh].xsize,
&v_spans[oh].center);
v_dbl_weights[oh].resize(interp_height);
ComputeAAWeightsCPU<WT>(
v_dbl_weights[oh].data(),
scale_h,
interp_height,
filter,
static_cast<WT>(v_spans[oh].xmin) - v_spans[oh].center,
v_spans[oh].xsize);
compute_int16_weights(v_dbl_weights[oh],
v_spans[oh].xsize,
v_i16_weights[oh],
v_precision[oh]);
}
// Temp buffer [N, H_in, W_out, C] as uint8
std::vector<uint8_t> temp(static_cast<size_t>(n) * in_h * out_w * c);
// Pass 1: Horizontal
for (int64_t bi = 0; bi < n; bi++) {
for (int ih = 0; ih < in_h; ih++) {
for (int ow = 0; ow < out_w; ow++) {
int xmin = h_spans[ow].xmin;
int xsize = h_spans[ow].xsize;
unsigned int P = h_precision[ow];
const int16_t* i16w = h_i16_weights[ow].data();
for (int64_t ch = 0; ch < c; ch++) {
int32_t accum = 1 << (P > 0 ? P - 1 : 0);
for (int j = 0; j < xsize; j++) {
int64_t in_idx = ((bi * in_h + ih) * in_w + (xmin + j)) * c + ch;
accum += static_cast<int32_t>(input_data[in_idx]) *
static_cast<int32_t>(i16w[j]);
}
int32_t result = accum >> P;
int64_t temp_idx = ((bi * in_h + ih) * out_w + ow) * c + ch;
temp[temp_idx] =
static_cast<uint8_t>(std::max(0, std::min(255, result)));
}
}
}
}
// Pass 2: Vertical
for (int64_t bi = 0; bi < n; bi++) {
for (int oh = 0; oh < out_h; oh++) {
int ymin = v_spans[oh].xmin;
int ysize = v_spans[oh].xsize;
unsigned int P = v_precision[oh];
const int16_t* i16w = v_i16_weights[oh].data();
for (int ow = 0; ow < out_w; ow++) {
for (int64_t ch = 0; ch < c; ch++) {
int32_t accum = 1 << (P > 0 ? P - 1 : 0);
for (int j = 0; j < ysize; j++) {
int64_t temp_idx = ((bi * in_h + (ymin + j)) * out_w + ow) * c + ch;
accum += static_cast<int32_t>(temp[temp_idx]) *
static_cast<int32_t>(i16w[j]);
}
int32_t result = accum >> P;
int64_t out_idx = ((bi * out_h + oh) * out_w + ow) * c + ch;
output_data[out_idx] =
static_cast<uint8_t>(std::max(0, std::min(255, result)));
}
}
}
}
}
// Dispatcher: selects NCHW/NHWC and float/uint8 paths
template <typename T, typename InterpFilter>
static void AAInterpolation2DCPUDispatch(const T* input_data,
T* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const DataLayout data_layout,
const InterpFilter& filter) {
if (data_layout == DataLayout::NCHW) {
AAInterpolation2DCPU_NCHW<T>(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
ratio_h,
ratio_w,
filter);
} else {
AAInterpolation2DCPU_NHWC<T>(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
ratio_h,
ratio_w,
filter);
}
}
// Explicit specialization for uint8_t dispatch
template <typename InterpFilter>
static void AAInterpolation2DCPUDispatchUInt8(const uint8_t* input_data,
uint8_t* output_data,
int64_t n,
int64_t c,
int in_h,
int in_w,
int out_h,
int out_w,
float ratio_h,
float ratio_w,
const DataLayout data_layout,
const InterpFilter& filter) {
if (data_layout == DataLayout::NCHW) {
AAInterpolation2DCPU_NCHW_UInt8(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
ratio_h,
ratio_w,
filter);
} else {
AAInterpolation2DCPU_NHWC_UInt8(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
ratio_h,
ratio_w,
filter);
}
}
// Main CPU forward function for AA 2D interpolation.
// Parses output size from out_size/size_tensor/scale_tensor/scale params
// (same logic as GPU InterpolateAA2DCUDAFwd), then dispatches to the
// separable 2-pass interpolation.
template <typename T, typename Context>
static void InterpolateAA2DCPUFwd(
const Context& dev_ctx,
const DenseTensor& input,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout_str,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
if (input.numel() == 0) {
dev_ctx.template Alloc<T>(output);
return;
}
auto* input_data = input.data<T>();
const DataLayout data_layout = StringToDataLayout(data_layout_str);
int64_t n, c, in_d, in_h, in_w;
funcs::ExtractNCDWH(input.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
double scale_w = -1;
double scale_h = -1;
if (size_tensor && !size_tensor->empty()) {
auto new_size = funcs::get_new_shape(size_tensor.get());
out_h = new_size[0];
out_w = new_size[1];
} else {
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
} else {
if (scale.size() > 1) {
scale_w = scale[1];
scale_h = scale[0];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
}
}
if (scale_w > 0. && scale_h > 0.) {
out_h = static_cast<int>(in_h * scale_h);
out_w = static_cast<int>(in_w * scale_w);
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
PADDLE_ENFORCE_GT(
out_h,
0,
errors::InvalidArgument("out_h in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
errors::InvalidArgument("out_w in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
DDim dim_out;
if (data_layout == DataLayout::NCHW) {
dim_out = {n, c, out_h, out_w};
} else {
dim_out = {n, out_h, out_w, c};
}
output->Resize(dim_out);
auto output_data = dev_ctx.template Alloc<T>(output);
if (in_h == out_h && in_w == out_w) {
Copy(dev_ctx, input, dev_ctx.GetPlace(), false, output);
return;
}
// Use conditional type: float for integral/half types, double for double
using MT = typename std::conditional_t<std::is_integral<T>::value,
float,
typename MPTypeTrait<T>::Type>;
MT ratio_h =
funcs::AreaPixelComputeScale<MT>(in_h, out_h, align_corners, scale_h);
MT ratio_w =
funcs::AreaPixelComputeScale<MT>(in_w, out_w, align_corners, scale_w);
// Dispatch based on interp_method and dtype
auto launch_aa = [&](auto filter_functor) {
if constexpr (std::is_same<T, uint8_t>::value) {
AAInterpolation2DCPUDispatchUInt8(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
static_cast<float>(ratio_h),
static_cast<float>(ratio_w),
data_layout,
filter_functor);
} else {
AAInterpolation2DCPUDispatch<T>(input_data,
output_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
static_cast<float>(ratio_h),
static_cast<float>(ratio_w),
data_layout,
filter_functor);
}
};
if ("bilinear" == interp_method) {
launch_aa(funcs::antialias::BilinearFilterFunctor{});
} else if ("bicubic" == interp_method) {
launch_aa(funcs::antialias::BicubicFilterFunctor{});
}
}
template <typename T, typename Context>
void InterpAntialiasKernel(
const Context& dev_ctx,
const DenseTensor& x,
const optional<DenseTensor>& out_size,
const optional<std::vector<const DenseTensor*>>& size_tensor,
const optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<double>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateAA2DCPUFwd<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
} // namespace phi
PD_REGISTER_KERNEL(bilinear_interp,
CPU,
ALL_LAYOUT,
phi::BilinearInterpKernel,
float,
double,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(legacy_bilinear_interp,
CPU,
ALL_LAYOUT,
phi::LegacyBilinearInterpKernel,
float,
double,
int,
int64_t,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(nearest_interp,
CPU,
ALL_LAYOUT,
phi::NearestInterpKernel,
float,
double,
int,
int64_t,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(legacy_nearest_interp,
CPU,
ALL_LAYOUT,
phi::LegacyNearestInterpKernel,
float,
double,
int,
int64_t,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(trilinear_interp,
CPU,
ALL_LAYOUT,
phi::TrilinearInterpKernel,
float,
double,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(linear_interp,
CPU,
ALL_LAYOUT,
phi::LinearInterpKernel,
float,
double,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(bicubic_interp,
CPU,
ALL_LAYOUT,
phi::BicubicInterpKernel,
float,
double,
phi::float16,
phi::bfloat16) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(interp_antialias,
CPU,
ALL_LAYOUT,
phi::InterpAntialiasKernel,
float,
double,
uint8_t,
phi::float16,
phi::bfloat16) {
kernel->InputAt(1).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}