I'm trying to understand how numpy can be so fast, based on my shocking comparison with optimized C/C++ code which is still far from reproducing numpy's speed.
Consider the following example:
Given a 2D array with shape=(N, N) and dtype=float32, which represents a list of N vectors of N dimensions, I am computing the pairwise differences between every pair of vectors. Using numpy broadcasting, this simply writes as:
def pairwise_sub_numpy( X ):
return X - X[:, None, :]
Using timeit I can measure the performance for N=512: it takes 88 ms per call on my laptop.
Now, in C/C++ a naive implementation writes as:
#define X(i, j) _X[(i)*N + (j)]
#define res(i, j, k) _res[((i)*N + (j))*N + (k)]
float* pairwise_sub_naive( const float* _X, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
for (int i = 0; i < N; i++) {
for (int j = 0; j < N; j++) {
for (int k = 0; k < N; k++)
res(i,j,k) = X(i,k) - X(j,k);
}
}
return _res;
}
Compiling using gcc 7.3.0 with -O3 flag, I get 195 ms per call for pairwise_sub_naive(X), which is not too bad given the simplicity of the code, but about 2 times slower than numpy.
Now I start getting serious and add some small optimizations, by indexing the row vectors directly:
float* pairwise_sub_better( const float* _X, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
for (int i = 0; i < N; i++) {
const float* xi = & X(i,0);
for (int j = 0; j < N; j++) {
const float* xj = & X(j,0);
float* r = &res(i,j,0);
for (int k = 0; k < N; k++)
r[k] = xi[k] - xj[k];
}
}
return _res;
}
The speed stays the same at 195 ms, which means that the compiler was able to figure that much. Let's now use SIMD vector instructions:
float* pairwise_sub_simd( const float* _X, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
// create caches for row vectors which are memory-aligned
float* xi = (float*)aligned_alloc(32, N * sizeof(float));
float* xj = (float*)aligned_alloc(32, N * sizeof(float));
for (int i = 0; i < N; i++) {
memcpy(xi, & X(i,0), N*sizeof(float));
for (int j = 0; j < N; j++) {
memcpy(xj, & X(j,0), N*sizeof(float));
float* r = &res(i,j,0);
for (int k = 0; k < N; k += 256/sizeof(float)) {
const __m256 A = _mm256_load_ps(xi+k);
const __m256 B = _mm256_load_ps(xj+k);
_mm256_store_ps(r+k, _mm256_sub_ps( A, B ));
}
}
}
free(xi);
free(xj);
return _res;
}
This only yields a small boost (178 ms instead of 194 ms per function call).
Then I was wondering if a "block-wise" approach, like what is used to optimize dot-products, could be beneficials:
float* pairwise_sub_blocks( const float* _X, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
#define B 8
float cache1[B*B], cache2[B*B];
for (int bi = 0; bi < N; bi+=B)
for (int bj = 0; bj < N; bj+=B)
for (int bk = 0; bk < N; bk+=B) {
// load first 8x8 block in the cache
for (int i = 0; i < B; i++)
for (int k = 0; k < B; k++)
cache1[B*i + k] = X(bi+i, bk+k);
// load second 8x8 block in the cache
for (int j = 0; j < B; j++)
for (int k = 0; k < B; k++)
cache2[B*j + k] = X(bj+j, bk+k);
// compute local operations on the caches
for (int i = 0; i < B; i++)
for (int j = 0; j < B; j++)
for (int k = 0; k < B; k++)
res(bi+i,bj+j,bk+k) = cache1[B*i + k] - cache2[B*j + k];
}
return _res;
}
And surprisingly, this is the slowest method so far (258 ms per function call).
To summarize, despite some efforts with some optimized C++ code, I can't come anywhere close the 88 ms / call that numpy achieves effortlessly. Any idea why?
Note: By the way, I am disabling numpy multi-threading and anyway, this kind of operation is not multi-threaded.
Edit: Exact code to benchmark the numpy code:
import numpy as np
def pairwise_sub_numpy( X ):
return X - X[:, None, :]
N = 512
X = np.random.rand(N,N).astype(np.float32)
import timeit
times = timeit.repeat('pairwise_sub_numpy( X )', globals=globals(), number=1, repeat=5)
print(f">> best of 5 = {1000*min(times):.3f} ms")
Full benchmark for C code:
#include <stdio.h>
#include <string.h>
#include <xmmintrin.h> // compile with -mavx -msse4.1
#include <pmmintrin.h>
#include <immintrin.h>
#include <time.h>
#define X(i, j) _x[(i)*N + (j)]
#define res(i, j, k) _res[((i)*N + (j))*N + (k)]
float* pairwise_sub_naive( const float* _x, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
for (int i = 0; i < N; i++) {
for (int j = 0; j < N; j++) {
for (int k = 0; k < N; k++)
res(i,j,k) = X(i,k) - X(j,k);
}
}
return _res;
}
float* pairwise_sub_better( const float* _x, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
for (int i = 0; i < N; i++) {
const float* xi = & X(i,0);
for (int j = 0; j < N; j++) {
const float* xj = & X(j,0);
float* r = &res(i,j,0);
for (int k = 0; k < N; k++)
r[k] = xi[k] - xj[k];
}
}
return _res;
}
float* pairwise_sub_simd( const float* _x, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
// create caches for row vectors which are memory-aligned
float* xi = (float*)aligned_alloc(32, N * sizeof(float));
float* xj = (float*)aligned_alloc(32, N * sizeof(float));
for (int i = 0; i < N; i++) {
memcpy(xi, & X(i,0), N*sizeof(float));
for (int j = 0; j < N; j++) {
memcpy(xj, & X(j,0), N*sizeof(float));
float* r = &res(i,j,0);
for (int k = 0; k < N; k += 256/sizeof(float)) {
const __m256 A = _mm256_load_ps(xi+k);
const __m256 B = _mm256_load_ps(xj+k);
_mm256_store_ps(r+k, _mm256_sub_ps( A, B ));
}
}
}
free(xi);
free(xj);
return _res;
}
float* pairwise_sub_blocks( const float* _x, int N )
{
float* _res = (float*) aligned_alloc( 32, N*N*N*sizeof(float));
#define B 8
float cache1[B*B], cache2[B*B];
for (int bi = 0; bi < N; bi+=B)
for (int bj = 0; bj < N; bj+=B)
for (int bk = 0; bk < N; bk+=B) {
// load first 8x8 block in the cache
for (int i = 0; i < B; i++)
for (int k = 0; k < B; k++)
cache1[B*i + k] = X(bi+i, bk+k);
// load second 8x8 block in the cache
for (int j = 0; j < B; j++)
for (int k = 0; k < B; k++)
cache2[B*j + k] = X(bj+j, bk+k);
// compute local operations on the caches
for (int i = 0; i < B; i++)
for (int j = 0; j < B; j++)
for (int k = 0; k < B; k++)
res(bi+i,bj+j,bk+k) = cache1[B*i + k] - cache2[B*j + k];
}
return _res;
}
int main()
{
const int N = 512;
float* _x = (float*) malloc( N * N * sizeof(float) );
for( int i = 0; i < N; i++)
for( int j = 0; j < N; j++)
X(i,j) = ((i+j*j+17*i+101) % N) / float(N);
double best = 9e9;
for( int i = 0; i < 5; i++)
{
struct timespec start, stop;
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start);
//float* res = pairwise_sub_naive( _x, N );
//float* res = pairwise_sub_better( _x, N );
//float* res = pairwise_sub_simd( _x, N );
float* res = pairwise_sub_blocks( _x, N );
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &stop);
double t = (stop.tv_sec - start.tv_sec) * 1e6 + (stop.tv_nsec - start.tv_nsec) / 1e3; // in microseconds
if (t < best) best = t;
free( res );
}
printf("Best of 5 = %f ms\n", best / 1000);
free( _x );
return 0;
}
Compiled using gcc 7.3.0 gcc -Wall -O3 -mavx -msse4.1 -o test_simd test_simd.c
Summary of timings on my machine:
Implementation
Time
numpy
88 ms
C++ naive
194 ms
C++ better
195 ms
C++ SIMD
178 ms
C++ blocked
258 ms
C++ blocked (gcc 8.3.1)
217 ms
As pointed out by some of the comments numpy uses SIMD in its implementation and it does not allocate memory at the point of computation. If I eliminate the memory allocation from your implementation, pre-allocating all the buffers ahead of the computation then I get a better time compared to numpy even with the scaler version(that is the one without any optimizations).
Also in terms of SIMD and why your implementation does not perform much better than the scaler is because your memory access patterns are not ideal for SIMD usage - you do memcopy and you load into SIMD registers from locations that are far apart from each other - e.g. you fill vectors from line 0 and line 511, which might not play well with the cache or with the SIMD prefetcher.
There is also a mistake in how you load the SIMD registers(if I understood correctly what you're trying to compute): a 256 bit SIMD register can load 8 single-precision floating-point numbers 8 * 32 = 256, but in your loop you jump k by "256/sizeof(float)" which is 256/4 = 64; _x and _res are float pointers and the SIMD intrinsics expect also float pointers as arguments so instead of reading all elements from those lines every 8 floats you read them every 64 floats.
The computation can be optimized further by changing the access patterns but also by observing that you repeat some computations: e.g. when iterating with line0 as a base you compute line0 - line1 but at some future time, when iterating with line1 as a base, you need to compute line1 - line0 which is basically -(line0 - line1), that is for each line after line0 a lot of results could be reused from previous computations.
A lot of times SIMD usage or parallelization requires one to change how data is accessed or reasoned about in order to provide meaningful improvements.
Here is what I have done as a first step based on your initial implementation and it is faster than the numpy(don't mind the OpenMP stuff as it's not how its supposed to be done, I just wanted to see how it behaves trying the naive way).
C++
Time scaler version: 55 ms
Time SIMD version: 53 ms
**Time SIMD 2 version: 33 ms**
Time SIMD 3 version: 168 ms
Time OpenMP version: 59 ms
Python numpy
>> best of 5 = 88.794 ms
#include <cstdlib>
#include <xmmintrin.h> // compile with -mavx -msse4.1
#include <pmmintrin.h>
#include <immintrin.h>
#include <numeric>
#include <algorithm>
#include <chrono>
#include <iostream>
#include <cstring>
using namespace std;
float* pairwise_sub_naive (const float* input, float* output, int n)
{
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
for (int k = 0; k < n; k++)
output[(i * n + j) * n + k] = input[i * n + k] - input[j * n + k];
}
}
return output;
}
float* pairwise_sub_simd (const float* input, float* output, int n)
{
for (int i = 0; i < n; i++)
{
const int idxi = i * n;
for (int j = 0; j < n; j++)
{
const int idxj = j * n;
const int outidx = idxi + j;
for (int k = 0; k < n; k += 8)
{
__m256 A = _mm256_load_ps(input + idxi + k);
__m256 B = _mm256_load_ps(input + idxj + k);
_mm256_store_ps(output + outidx * n + k, _mm256_sub_ps( A, B ));
}
}
}
return output;
}
float* pairwise_sub_simd_2 (const float* input, float* output, int n)
{
float* line_buffer = (float*) aligned_alloc(32, n * sizeof(float));
for (int i = 0; i < n; i++)
{
const int idxi = i * n;
for (int j = 0; j < n; j++)
{
const int idxj = j * n;
const int outidx = idxi + j;
for (int k = 0; k < n; k += 8)
{
__m256 A = _mm256_load_ps(input + idxi + k);
__m256 B = _mm256_load_ps(input + idxj + k);
_mm256_store_ps(line_buffer + k, _mm256_sub_ps( A, B ));
}
memcpy(output + outidx * n, line_buffer, n);
}
}
return output;
}
float* pairwise_sub_simd_3 (const float* input, float* output, int n)
{
for (int i = 0; i < n; i++)
{
const int idxi = i * n;
for (int k = 0; k < n; k += 8)
{
__m256 A = _mm256_load_ps(input + idxi + k);
for (int j = 0; j < n; j++)
{
const int idxj = j * n;
const int outidx = (idxi + j) * n;
__m256 B = _mm256_load_ps(input + idxj + k);
_mm256_store_ps(output + outidx + k, _mm256_sub_ps( A, B ));
}
}
}
return output;
}
float* pairwise_sub_openmp (const float* input, float* output, int n)
{
int i, j;
#pragma omp parallel for private(j)
for (i = 0; i < n; i++)
{
for (j = 0; j < n; j++)
{
const int idxi = i * n;
const int idxj = j * n;
const int outidx = idxi + j;
for (int k = 0; k < n; k += 8)
{
__m256 A = _mm256_load_ps(input + idxi + k);
__m256 B = _mm256_load_ps(input + idxj + k);
_mm256_store_ps(output + outidx * n + k, _mm256_sub_ps( A, B ));
}
}
}
/*for (i = 0; i < n; i++)
{
for (j = 0; j < n; j++)
{
for (int k = 0; k < n; k++)
{
output[(i * n + j) * n + k] = input[i * n + k] - input[j * n + k];
}
}
}*/
return output;
}
int main ()
{
constexpr size_t n = 512;
constexpr size_t input_size = n * n;
constexpr size_t output_size = n * n * n;
float* input = (float*) aligned_alloc(32, input_size * sizeof(float));
float* output = (float*) aligned_alloc(32, output_size * sizeof(float));
float* input_simd = (float*) aligned_alloc(32, input_size * sizeof(float));
float* output_simd = (float*) aligned_alloc(32, output_size * sizeof(float));
float* input_par = (float*) aligned_alloc(32, input_size * sizeof(float));
float* output_par = (float*) aligned_alloc(32, output_size * sizeof(float));
iota(input, input + input_size, float(0.0));
fill(output, output + output_size, float(0.0));
iota(input_simd, input_simd + input_size, float(0.0));
fill(output_simd, output_simd + output_size, float(0.0));
iota(input_par, input_par + input_size, float(0.0));
fill(output_par, output_par + output_size, float(0.0));
std::chrono::milliseconds best_scaler{100000};
for (int i = 0; i < 5; ++i)
{
auto start = chrono::high_resolution_clock::now();
pairwise_sub_naive(input, output, n);
auto stop = chrono::high_resolution_clock::now();
auto duration = chrono::duration_cast<chrono::milliseconds>(stop - start);
if (duration < best_scaler)
{
best_scaler = duration;
}
}
cout << "Time scaler version: " << best_scaler.count() << " ms\n";
std::chrono::milliseconds best_simd{100000};
for (int i = 0; i < 5; ++i)
{
auto start = chrono::high_resolution_clock::now();
pairwise_sub_simd(input_simd, output_simd, n);
auto stop = chrono::high_resolution_clock::now();
auto duration = chrono::duration_cast<chrono::milliseconds>(stop - start);
if (duration < best_simd)
{
best_simd = duration;
}
}
cout << "Time SIMD version: " << best_simd.count() << " ms\n";
std::chrono::milliseconds best_simd_2{100000};
for (int i = 0; i < 5; ++i)
{
auto start = chrono::high_resolution_clock::now();
pairwise_sub_simd_2(input_simd, output_simd, n);
auto stop = chrono::high_resolution_clock::now();
auto duration = chrono::duration_cast<chrono::milliseconds>(stop - start);
if (duration < best_simd_2)
{
best_simd_2 = duration;
}
}
cout << "Time SIMD 2 version: " << best_simd_2.count() << " ms\n";
std::chrono::milliseconds best_simd_3{100000};
for (int i = 0; i < 5; ++i)
{
auto start = chrono::high_resolution_clock::now();
pairwise_sub_simd_3(input_simd, output_simd, n);
auto stop = chrono::high_resolution_clock::now();
auto duration = chrono::duration_cast<chrono::milliseconds>(stop - start);
if (duration < best_simd_3)
{
best_simd_3 = duration;
}
}
cout << "Time SIMD 3 version: " << best_simd_3.count() << " ms\n";
std::chrono::milliseconds best_par{100000};
for (int i = 0; i < 5; ++i)
{
auto start = chrono::high_resolution_clock::now();
pairwise_sub_openmp(input_par, output_par, n);
auto stop = chrono::high_resolution_clock::now();
auto duration = chrono::duration_cast<chrono::milliseconds>(stop - start);
if (duration < best_par)
{
best_par = duration;
}
}
cout << "Time OpenMP version: " << best_par.count() << " ms\n";
cout << "Verification\n";
if (equal(output, output + output_size, output_simd))
{
cout << "PASSED\n";
}
else
{
cout << "FAILED\n";
}
return 0;
}
Edit: Small correction as there was a wrong call related to the second version of SIMD implementation.
As you can see now, the second implementation is the fastest as it behaves the best from the point of view of the locality of reference of the cache. Examples 2 and 3 of SIMD implementations are there to illustrate for you how changing memory access patterns to influence the performance of your SIMD optimizations.
To summarize(knowing that I'm far from being complete in my advice) be mindful of your memory access patterns and of the loads and stores to\from the SIMD unit; the SIMD is a different hardware unit inside the processor's core so there is a penalty in shuffling data back and forth, hence when you load a register from memory try to do as many operations as possible with that data and do not be too eager to store it back(of course, in your example that might be all you need to do with the data). Be mindful also that there is a limited number of SIMD registers available and if you load too many then they will "spill", that is they will be stored back to temporary locations in main memory behind the scenes killing all your gains. SIMD optimization, it's a true balance act!
There is some effort to put a cross-platform intrinsics wrapper into the standard(I developed myself a closed source one in my glorious past) and even it's far from being complete, it's worth taking a look at(read the accompanying papers if you're truly interested to learn how SIMD works).
https://github.com/VcDevel/std-simd
This is a complement to the answer posted by #celakev .
I think I finally got to understand what exactly was the issue. The issue was not about allocating the memory in the main function that does the computation.
What was actually taking time is to access new (fresh) memory. I believe that the malloc call returns pages of memory which are virtual, i.e. that does not corresponds to actual physical memory -- until it is explicitly accessed. What actually takes time is the process of allocating physical memory on the fly (which I think is OS-level) when it is accessed in the function code.
Here is a proof. Consider the two following trivial functions:
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
float* just_alloc( size_t N )
{
return (float*) aligned_alloc( 32, sizeof(float)*N );
}
void just_fill( float* _arr, size_t N )
{
for (size_t i = 0; i < N; i++)
_arr[i] = 1;
}
#define Time( code_to_benchmark, cleanup_code ) \
do { \
double best = 9e9; \
for( int i = 0; i < 5; i++) { \
struct timespec start, stop; \
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start); \
code_to_benchmark; \
clock_gettime(CLOCK_THREAD_CPUTIME_ID, &stop); \
double t = (stop.tv_sec - start.tv_sec) * 1e3 + (stop.tv_nsec - start.tv_nsec) / 1e6; \
printf("Time[%d] = %f ms\n", i, t); \
if (t < best) best = t; \
cleanup_code; \
} \
printf("Best of 5 for '" #code_to_benchmark "' = %f ms\n\n", best); \
} while(0)
int main()
{
const size_t N = 512;
Time( float* arr = just_alloc(N*N*N), free(arr) );
float* arr = just_alloc(N*N*N);
Time( just_fill(arr, N*N*N), ; );
free(arr);
return 0;
}
I get the following timings, which I now detail for each of the calls:
Time[0] = 0.000931 ms
Time[1] = 0.000540 ms
Time[2] = 0.000523 ms
Time[3] = 0.000524 ms
Time[4] = 0.000521 ms
Best of 5 for 'float* arr = just_alloc(N*N*N)' = 0.000521 ms
Time[0] = 189.822237 ms
Time[1] = 45.041083 ms
Time[2] = 46.331428 ms
Time[3] = 44.729433 ms
Time[4] = 42.241279 ms
Best of 5 for 'just_fill(arr, N*N*N)' = 42.241279 ms
As you can see, allocating memory is blazingly fast, but the first time that the memory is accessed, it is 5 times slower than the other times. So, basically the reason that my code was slow was because i was each time reallocating fresh memory that had no physical address yet. (Correct me if I'm wrong but I think that's the gist of it!)
A bit late to the party, but I wanted to add a pairwise method with Eigen, which is supposed to give C++ a high-level algebra manipulation capability and use SIMD under the hood. Just like numpy.
Here is the implementation
#include <iostream>
#include <vector>
#include <chrono>
#include <algorithm>
#include <Eigen/Dense>
auto pairwise_eigen(const Eigen::MatrixXf &input, std::vector<Eigen::MatrixXf> &output) {
for (int k = 0; k < input.cols(); ++k)
output[k] = input
// subtract matrix with repeated k-th column
- input.col(k) * Eigen::RowVectorXf::Ones(input.cols());
}
int main() {
constexpr size_t n = 512;
// allocate input and output
Eigen::MatrixXf input = Eigen::MatrixXf::Random(n, n);
std::vector<Eigen::MatrixXf> output(n);
std::chrono::milliseconds best_eigen{100000};
for (int i = 0; i < 5; ++i) {
auto start = std::chrono::high_resolution_clock::now();
pairwise_eigen(input, output);
auto end = std::chrono::high_resolution_clock::now();
auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end-start);
if (duration < best_eigen)
best_eigen = duration;
}
std::cout << "Time Eigen version: " << best_eigen.count() << " ms\n";
return 0;
}
The full benchmark tests suggested by #celavek on my system are
Time scaler version: 57 ms
Time SIMD version: 58 ms
Time SIMD 2 version: 40 ms
Time SIMD 3 version: 58 ms
Time OpenMP version: 58 ms
Time Eigen version: 76 ms
Numpy >> best of 5 = 118.489 ms
Whit Eigen there is still a noticeable improvement with respect to Numpy, but not so impressive compared to the "raw" implementations (there is certainly some overhead).
An extra optimization is to allocate the output vector with copies of the input and then subtract directly from each vector entry, simply replacing the following lines
// inside the pairwise method
for (int k = 0; k < input.cols(); ++k)
output[k] -= input.col(k) * Eigen::RowVectorXf::Ones(input.cols());
// at allocation time
std::vector<Eigen::MatrixXf> output(n, input);
This pushes the best of 5 down to 60 ms.
I am working on a raytracer. I wanted to optimize my code by saving data of each pixel in an OpenCV Mat, using GPU.
For now, I save pixel values in buffer fb which is a vector of three values (RGB):
__global__ void render(vec3 *fb, int max_x, int max_y, Camera **cam, Triangle *data, size_t n, )
{
int i = threadIdx.x + blockIdx.x * blockDim.x;
int j = threadIdx.y + blockIdx.y * blockDim.y;
if ((i >= max_x) || (j >= max_y)) return;
int pixel_index = j * max_x + i;
float u = float(i) / float(max_x);
float v = float(j) / float(max_y);
Ray r = (*cam)->get_ray(u,v);
fb[pixel_index] = color(r, data,n);
}
and then I save data in Mat on CPU:
for (int j = ny - 1; j >= 0; j--)
{
for (int i = 0; i < nx; i++)
{
size_t pixel_index = j * nx + i;
int ir = int(255.99*fb[pixel_index].r());
int ig = int(255.99*fb[pixel_index].g());
int ib = int(255.99*fb[pixel_index].b());
output.at<Vec3b>(j, i)[0] = (uchar)ib;
output.at<Vec3b>(j, i)[1] = (uchar)ig;
output.at<Vec3b>(j, i)[2] = (uchar)ir;
//std::cout << ir << " " << ig << " " << ib << "\n";
}
}
but it is a very slow process when I have a large pixel array. That why I want to use an OpenCV GpuMat and save the data directly on the GPU.
The problem is that I can't really find an example of how I can save data in each channel of GPU Mat. Is it an easy way to do it, similar to saving data on the CPU?
See the documentation. There it says
no functions that return references to their data (because references on GPU are not valid for CPU)
The only way to access the data is through the datafunction. But the pointer can only be dereferenced in (cuda)kernel code. And there is no
at function as far as I see. So you will have to calculate the offset from data.
Thank you for your answers. They make me think how to do it in another way. I am not sure if it is the best solution to deal with it but it works and in my opinion is a quiet easy way to fill matrix on GPU.
Reserve memory on GPU for matrix
Mat output(ny, nx, CV_8UC3);
const size_t numBytes = output.step * output.rows;
unsigned char *d_output;
cudaMalloc<unsigned char>(&d_output, numBytes);
Fill matrix on GPU
_global__ void render(vec3 *fb, int max_x, int max_y, Camera **cam, Triangle *data, size_t n, unsigned char* input, int step)
{
int i = threadIdx.x + blockIdx.x * blockDim.x;
int j = threadIdx.y + blockIdx.y * blockDim.y;
if ((i >= max_x) || (j >= max_y)) return;
int pixel_index = j * max_x + i;
int index = j * step + 3 * i;
float u = float(i) / float(max_x);
float v = float(j) / float(max_y);
Ray r = (*cam)->get_ray(u,v);
fb[pixel_index] = color(r, data,n);
int ir = int(255.99*fb[pixel_index].r());
int ig = int(255.99*fb[pixel_index].g());
int ib = int(255.99*fb[pixel_index].b());
input[index] = ib;
input[index+1] = ig;
input[index+2] = ir;
}
I will be grateful for any advice and comments to this code.
everyone I am trying to implement patter matching with FFT but I am not sure what the result should be (I think I am missing something even though a read a lot of stuff about the problem and tried a lot of different implementations this one is the best so far). Here is my FFT correlation function.
void fft2d(fftw_complex**& a, int rows, int cols, bool forward = true)
{
fftw_plan p;
for (int i = 0; i < rows; ++i)
{
p = fftw_plan_dft_1d(cols, a[i], a[i], forward ? FFTW_FORWARD : FFTW_BACKWARD, FFTW_ESTIMATE);
fftw_execute(p);
}
fftw_complex* t = (fftw_complex*)fftw_malloc(rows * sizeof(fftw_complex));
for (int j = 0; j < cols; ++j)
{
for (int i = 0; i < rows; ++i)
{
t[i][0] = a[i][j][0];
t[i][1] = a[i][j][1];
}
p = fftw_plan_dft_1d(rows, t, t, forward ? FFTW_FORWARD : FFTW_BACKWARD, FFTW_ESTIMATE);
fftw_execute(p);
for (int i = 0; i < rows; ++i)
{
a[i][j][0] = t[i][0];
a[i][j][1] = t[i][1];
}
}
fftw_free(t);
}
int findCorrelation(int argc, char* argv[])
{
BMP bigImage;
BMP keyImage;
BMP result;
RGBApixel blackPixel = { 0, 0, 0, 1 };
const bool swapQuadrants = (argc == 4);
if (argc < 3 || argc > 4) {
cout << "correlation img1.bmp img2.bmp" << endl;
return 1;
}
if (!keyImage.ReadFromFile(argv[1])) {
return 1;
}
if (!bigImage.ReadFromFile(argv[2])) {
return 1;
}
//Preparations
const int maxWidth = std::max(bigImage.TellWidth(), keyImage.TellWidth());
const int maxHeight = std::max(bigImage.TellHeight(), keyImage.TellHeight());
const int rowsCount = maxHeight;
const int colsCount = maxWidth;
BMP bigTemp = bigImage;
BMP keyTemp = keyImage;
keyImage.SetSize(maxWidth, maxHeight);
bigImage.SetSize(maxWidth, maxHeight);
for (int i = 0; i < rowsCount; ++i)
for (int j = 0; j < colsCount; ++j) {
RGBApixel p1;
if (i < bigTemp.TellHeight() && j < bigTemp.TellWidth()) {
p1 = bigTemp.GetPixel(j, i);
} else {
p1 = blackPixel;
}
bigImage.SetPixel(j, i, p1);
RGBApixel p2;
if (i < keyTemp.TellHeight() && j < keyTemp.TellWidth()) {
p2 = keyTemp.GetPixel(j, i);
} else {
p2 = blackPixel;
}
keyImage.SetPixel(j, i, p2);
}
//Here is where the transforms begin
fftw_complex **a = (fftw_complex**)fftw_malloc(rowsCount * sizeof(fftw_complex*));
fftw_complex **b = (fftw_complex**)fftw_malloc(rowsCount * sizeof(fftw_complex*));
fftw_complex **c = (fftw_complex**)fftw_malloc(rowsCount * sizeof(fftw_complex*));
for (int i = 0; i < rowsCount; ++i) {
a[i] = (fftw_complex*)fftw_malloc(colsCount * sizeof(fftw_complex));
b[i] = (fftw_complex*)fftw_malloc(colsCount * sizeof(fftw_complex));
c[i] = (fftw_complex*)fftw_malloc(colsCount * sizeof(fftw_complex));
for (int j = 0; j < colsCount; ++j) {
RGBApixel p1;
p1 = bigImage.GetPixel(j, i);
a[i][j][0] = (0.299*p1.Red + 0.587*p1.Green + 0.114*p1.Blue);
a[i][j][1] = 0.0;
RGBApixel p2;
p2 = keyImage.GetPixel(j, i);
b[i][j][0] = (0.299*p2.Red + 0.587*p2.Green + 0.114*p2.Blue);
b[i][j][1] = 0.0;
}
}
fft2d(a, rowsCount, colsCount);
fft2d(b, rowsCount, colsCount);
result.SetSize(maxWidth, maxHeight);
for (int i = 0; i < rowsCount; ++i)
for (int j = 0; j < colsCount; ++j) {
fftw_complex& y = a[i][j];
fftw_complex& x = b[i][j];
double u = x[0], v = x[1];
double m = y[0], n = y[1];
c[i][j][0] = u*m + n*v;
c[i][j][1] = v*m - u*n;
int fx = j;
if (fx>(colsCount / 2)) fx -= colsCount;
int fy = i;
if (fy>(rowsCount / 2)) fy -= rowsCount;
float r2 = (fx*fx + fy*fy);
const double cuttoffCoef = (maxWidth * maxHeight) / 37992.;
if (r2<128 * 128 * cuttoffCoef)
c[i][j][0] = c[i][j][1] = 0;
}
fft2d(c, rowsCount, colsCount, false);
const int halfCols = colsCount / 2;
const int halfRows = rowsCount / 2;
if (swapQuadrants) {
for (int i = 0; i < halfRows; ++i)
for (int j = 0; j < halfCols; ++j) {
std::swap(c[i][j][0], c[i + halfRows][j + halfCols][0]);
std::swap(c[i][j][1], c[i + halfRows][j + halfCols][1]);
}
for (int i = halfRows; i < rowsCount; ++i)
for (int j = 0; j < halfCols; ++j) {
std::swap(c[i][j][0], c[i - halfRows][j + halfCols][0]);
std::swap(c[i][j][1], c[i - halfRows][j + halfCols][1]);
}
}
for (int i = 0; i < rowsCount; ++i)
for (int j = 0; j < colsCount; ++j) {
const double& g = c[i][j][0];
RGBApixel pixel;
pixel.Alpha = 0;
int gInt = 255 - static_cast<int>(std::floor(g + 0.5));
pixel.Red = gInt;
pixel.Green = gInt;
pixel.Blue = gInt;
result.SetPixel(j, i, pixel);
}
BMP res;
res.SetSize(maxWidth, maxHeight);
result.WriteToFile("result.bmp");
return 0;
}
Sample output
This question would probably be more appropriately posted on another site like cross validated (metaoptimize.com used to also be a good one, but it appears to be gone)
That said:
There's two similar operations you can perform with FFT: convolution and correlation. Convolution is used for determining how two signals interact with each-other, whereas correlation can be used to express how similar two signals are to each-other. Make sure you're doing the right operation as they're both commonly implemented throught a DFT.
For this type of application of DFTs you usually wouldn't extract any useful information in the fourier spectrum unless you were looking for frequencies common to both data sources or whatever (eg, if you were comparing two bridges to see if their supports are spaced similarly).
Your 3rd image looks a lot like the power domain; normally I see the correlation output entirely grey except where overlap occurred. Your code definitely appears to be computing the inverse DFT, so unless I'm missing something the only other explanation I've come up with for the fuzzy look could be some of the "fudge factor" code in there like:
if (r2<128 * 128 * cuttoffCoef)
c[i][j][0] = c[i][j][1] = 0;
As for what you should expect: wherever there are common elements between the two images you'll see a peak. The larger the peak, the more similar the two images are near that region.
Some comments and/or recommended changes:
1) Convolution & correlation are not scale invariant operations. In other words, the size of your pattern image can make a significant difference in your output.
2) Normalize your images before correlation.
When you get the image data ready for the forward DFT pass:
a[i][j][0] = (0.299*p1.Red + 0.587*p1.Green + 0.114*p1.Blue);
a[i][j][1] = 0.0;
/* ... */
How you grayscale the image is your business (though I would've picked something like sqrt( r*r + b*b + g*g )). However, I don't see you doing anything to normalize the image.
The word "normalize" can take on a few different meanings in this context. Two common types:
normalize the range of values between 0.0 and 1.0
normalize the "whiteness" of the images
3) Run your pattern image through an edge enhancement filter. I've personally made use of canny, sobel, and I think I messed with a few others. As I recall, canny was "quick'n dirty", sobel was more expensive, but I got comparable results when it came time to do correlation. See chapter 24 of the "dsp guide" book that's freely available online. The whole book is worth your time, but if you're low on time then at a minimum chapter 24 will help a lot.
4) Re-scale the output image between [0, 255]; if you want to implement thresholds, do it after this step because the thresholding step is lossy.
My memory on this one is hazy, but as I recall (edited for clarity):
You can scale the final image pixels (before rescaling) between [-1.0, 1.0] by dividing off the largest power spectrum value from the entire power spectrum
The largest power spectrum value is, conveniently enough, the center-most value in the power spectrum (corresponding to the lowest frequency)
If you divide it off the power spectrum, you'll end up doing twice the work; since FFTs are linear, you can delay the division until after the inverse DFT pass to when you're re-scaling the pixels between [0..255].
If after rescaling most of your values end up so black you can't see them, you can use a solution to the ODE y' = y(1 - y) (one example is the sigmoid f(x) = 1 / (1 + exp(-c*x) ), for some scaling factor c that gives better gradations). This has more to do with improving your ability to interpret the results visually than anything you might use to programmatically find peaks.
edit I said [0, 255] above. I suggest you rescale to [128, 255] or some other lower bound that is gray rather than black.
I have a 3007 x 1644 dimensional matrix of terms and documents. I am trying to assign weights to frequency of terms in each document so I'm using this log entropy formula http://en.wikipedia.org/wiki/Latent_semantic_indexing#Term_Document_Matrix (See entropy formula in the last row).
I'm successfully doing this but my code is running for >7 minutes.
Here's the code:
int N = mat.cols();
for(int i=1;i<=mat.rows();i++){
double gfi = sum(mat(i,colon()))(1,1); //sum of occurrence of terms
double g =0;
if(gfi != 0){// to avoid divide by zero error
for(int j = 1;j<=N;j++){
double tfij = mat(i,j);
double pij = gfi==0?0.0:tfij/gfi;
pij = pij + 1; //avoid log0
double G = (pij * log(pij))/log(N);
g = g + G;
}
}
double gi = 1 - g;
for(int j=1;j<=N;j++){
double tfij = mat(i,j) + 1;//avoid log0
double aij = gi * log(tfij);
mat(i,j) = aij;
}
}
Anyone have ideas how I can optimize this to make it faster? Oh and mat is a RealSparseMatrix from amlpp matrix library.
UPDATE
Code runs on Linux mint with 4gb RAM and AMD Athlon II dual core
Running time before change: > 7mins
After #Kereks answer: 4.1sec
Here's a very naive rewrite that removes some redundancies:
int const N = mat.cols();
double const logN = log(N);
for (int i = 1; i <= mat.rows(); ++i)
{
double const gfi = sum(mat(i, colon()))(1, 1); // sum of occurrence of terms
double g = 0;
if (gfi != 0)
{
for (int j = 1; j <= N; ++j)
{
double const pij = mat(i, j) / gfi + 1;
g += pij * log(pij);
}
g /= logN;
}
for (int j = 1; j <= N; ++j)
{
mat(i,j) = (1 - g) * log(mat(i, j) + 1);
}
}
Also make sure that the matrix data structure is sane (e.g. a flat array accessed in strides; not a bunch of dynamically allocated rows).
Also, I think the first + 1 is a bit silly. You know that x -> x * log(x) is continuous at zero with limit zero, so you should write:
double const pij = mat(i, j) / gfi;
if (pij != 0) { g += pij + log(pij); }
In fact, you might even write the first inner for loop like this, avoiding a division when it isn't needed:
for (int j = 1; j <= N; ++j)
{
if (double pij = mat(i, j))
{
pij /= gfi;
g += pij * log(pij);
}
}