Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[pdq] Constant initialize DCT buffer #1518

Merged
merged 2 commits into from
Jan 22, 2024
Merged

[pdq] Constant initialize DCT buffer #1518

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b1d557b
Constant initialize DCT matrix instead of using std::call_once
ianwal Jan 21, 2024
a2df0f4
Add venv to .gitignore
ianwal Jan 21, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,4 +18,5 @@ pdq/java/bazel-testlogs/
.mypy_cache/
.vscode/
*.d.ts
.idea/
.idea/
venv/
69 changes: 35 additions & 34 deletions pdq/cpp/hashing/pdqhashing.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,6 @@
// Copyright (c) Meta Platforms, Inc. and affiliates.
// ================================================================

#include <mutex>

#include <pdq/cpp/downscaling/downscaling.h>
#include <pdq/cpp/hashing/pdqhashing.h>
#include <pdq/cpp/hashing/torben.h>
Expand All @@ -17,6 +15,7 @@
#define _USE_MATH_DEFINES
#endif

#include <array>
#include <cassert>
#include <chrono>
#include <cmath>
Expand All @@ -27,6 +26,37 @@ namespace facebook {
namespace pdq {
namespace hashing {

namespace {

// ----------------------------------------------------------------
// Christoph Zauner 'Implementation and Benchmarking of Perceptual
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For fun, I found this paper and tried to figure out what this was referring to, and I can't find anything on the page that seems to match.

// Image Hash Functions' 2010
//
// See comments on dct64To16. Input is (0..63)x(0..63); output is
// (1..16)x(1..16) with the latter indexed as (0..15)x(0..15).
//
// * numRows is 16.
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
auto const dct_matrix_64 = [] {
const size_t num_rows = 16;
const size_t num_cols = 64;
const float matrix_scale_factor = std::sqrt(2.0 / double{num_cols});

std::array<float, (num_rows * num_cols)> dct_matrix;
for (size_t i = 0; i < num_rows; i++) {
for (size_t j = 0; j < num_cols; j++) {
dct_matrix[i * num_cols + j] = matrix_scale_factor *
std::cos((M_PI / 2.0 / double{num_cols}) * (i + 1) * (2 * j + 1));
}
}

return dct_matrix;
}();

} // namespace

// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
// From Wikipedia: standard RGB to luminance (the 'Y' in 'YUV').
const float luma_from_R_coeff = 0.299;
Expand All @@ -41,11 +71,6 @@ const int MIN_HASHABLE_DIM = 5;
// Tent filter.
const int PDQ_NUM_JAROSZ_XY_PASSES = 2;

// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
// Christoph Zauner 'Implementation and Benchmarking of Perceptual
// Image Hash Functions' 2010
static float* fill_dct_matrix_64_cached();

// ----------------------------------------------------------------
void fillFloatLumaFromRGB(
uint8_t* pRbase,
Expand Down Expand Up @@ -333,15 +358,15 @@ void dct64To16(float A[64][64], float T[16][64], float B[16][16]) {
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
float* D = fill_dct_matrix_64_cached();
const auto& D = dct_matrix_64;

// B = D A Dt
// B = (D A) Dt
// with intermediate T = D A

for (int i = 0; i < 16; i++) {
for (int j = 0; j < 64; j++) {
float* pd = &D[i * 64]; // ith row
const auto pd = &D[i * 64]; // ith row
float* pa = &A[0][j];
float sumk = 0.0;

Expand Down Expand Up @@ -371,7 +396,7 @@ void dct64To16(float A[64][64], float T[16][64], float B[16][16]) {
for (int i = 0; i < 16; i++) {
for (int j = 0; j < 16; j++) {
float sumk = 0.0;
float* pd = &D[j * 64]; // jth row
const auto pd = &D[j * 64]; // jth row
float* pt = &T[i][0];
for (int k = 0; k < 64;) {
sumk += pt[k] * pd[k];
Expand Down Expand Up @@ -524,30 +549,6 @@ void pdqBuffer16x16ToBits(float dctOutput16x16[16][16], Hash256* hashptr) {
}
}

// ----------------------------------------------------------------
// See comments on dct64To16. Input is (0..63)x(0..63); output is
// (1..16)x(1..16) with the latter indexed as (0..15)x(0..15).
//
// * numRows is 16.
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
static float* fill_dct_matrix_64_cached() {
static std::once_flag initialized;
static float buffer[16 * 64];

std::call_once(initialized, []() {
const float matrix_scale_factor = std::sqrt(2.0 / 64.0);
for (int i = 0; i < 16; i++) {
for (int j = 0; j < 64; j++) {
buffer[i * 64 + j] = matrix_scale_factor *
cos((M_PI / 2 / 64.0) * (i + 1) * (2 * j + 1));
}
}
});
return &buffer[0];
}

} // namespace hashing
} // namespace pdq
} // namespace facebook
69 changes: 35 additions & 34 deletions vpdq/cpp/pdq/cpp/hashing/pdqhashing.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,6 @@
// Copyright (c) Meta Platforms, Inc. and affiliates.
// ================================================================

#include <mutex>
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any ideas for in the long term how we can ensure pdq and vpdq say in sync? Should we write a test that checks that file contents of the vpdq directory match the cpp one?


#include <pdq/cpp/downscaling/downscaling.h>
#include <pdq/cpp/hashing/pdqhashing.h>
#include <pdq/cpp/hashing/torben.h>
Expand All @@ -17,6 +15,7 @@
#define _USE_MATH_DEFINES
#endif

#include <array>
#include <cassert>
#include <chrono>
#include <cmath>
Expand All @@ -27,6 +26,37 @@ namespace facebook {
namespace pdq {
namespace hashing {

namespace {

// ----------------------------------------------------------------
// Christoph Zauner 'Implementation and Benchmarking of Perceptual
// Image Hash Functions' 2010
//
// See comments on dct64To16. Input is (0..63)x(0..63); output is
// (1..16)x(1..16) with the latter indexed as (0..15)x(0..15).
//
// * numRows is 16.
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
auto const dct_matrix_64 = [] {
const size_t num_rows = 16;
const size_t num_cols = 64;
const float matrix_scale_factor = std::sqrt(2.0 / double{num_cols});

std::array<float, (num_rows * num_cols)> dct_matrix;
for (size_t i = 0; i < num_rows; i++) {
for (size_t j = 0; j < num_cols; j++) {
dct_matrix[i * num_cols + j] = matrix_scale_factor *
std::cos((M_PI / 2.0 / double{num_cols}) * (i + 1) * (2 * j + 1));
}
}

return dct_matrix;
}();

} // namespace

// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
// From Wikipedia: standard RGB to luminance (the 'Y' in 'YUV').
const float luma_from_R_coeff = 0.299;
Expand All @@ -41,11 +71,6 @@ const int MIN_HASHABLE_DIM = 5;
// Tent filter.
const int PDQ_NUM_JAROSZ_XY_PASSES = 2;

// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
// Christoph Zauner 'Implementation and Benchmarking of Perceptual
// Image Hash Functions' 2010
static float* fill_dct_matrix_64_cached();

// ----------------------------------------------------------------
void fillFloatLumaFromRGB(
uint8_t* pRbase,
Expand Down Expand Up @@ -333,15 +358,15 @@ void dct64To16(float A[64][64], float T[16][64], float B[16][16]) {
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
float* D = fill_dct_matrix_64_cached();
const auto& D = dct_matrix_64;

// B = D A Dt
// B = (D A) Dt
// with intermediate T = D A

for (int i = 0; i < 16; i++) {
for (int j = 0; j < 64; j++) {
float* pd = &D[i * 64]; // ith row
const auto pd = &D[i * 64]; // ith row
float* pa = &A[0][j];
float sumk = 0.0;

Expand Down Expand Up @@ -371,7 +396,7 @@ void dct64To16(float A[64][64], float T[16][64], float B[16][16]) {
for (int i = 0; i < 16; i++) {
for (int j = 0; j < 16; j++) {
float sumk = 0.0;
float* pd = &D[j * 64]; // jth row
const auto pd = &D[j * 64]; // jth row
float* pt = &T[i][0];
for (int k = 0; k < 64;) {
sumk += pt[k] * pd[k];
Expand Down Expand Up @@ -524,30 +549,6 @@ void pdqBuffer16x16ToBits(float dctOutput16x16[16][16], Hash256* hashptr) {
}
}

// ----------------------------------------------------------------
// See comments on dct64To16. Input is (0..63)x(0..63); output is
// (1..16)x(1..16) with the latter indexed as (0..15)x(0..15).
//
// * numRows is 16.
// * numCols is 64.
// * Storage is row-major
// * Element i,j at row i column j is at offset i*16+j.
static float* fill_dct_matrix_64_cached() {
static std::once_flag initialized;
static float buffer[16 * 64];

std::call_once(initialized, []() {
const float matrix_scale_factor = std::sqrt(2.0 / 64.0);
for (int i = 0; i < 16; i++) {
for (int j = 0; j < 64; j++) {
buffer[i * 64 + j] = matrix_scale_factor *
cos((M_PI / 2 / 64.0) * (i + 1) * (2 * j + 1));
}
}
});
return &buffer[0];
}

} // namespace hashing
} // namespace pdq
} // namespace facebook
Loading