I am working on google colab and i want to use libsvm library in my project. I downloaded libsvm and installed it. Now when i use !nvcc -o command and run the code using CUDA i am getting errors like,
undefined reference to `svm_get_nr_class
undefined reference to 'svm_predict_probability'
undefined reference to `svm_free_and_destroy_model
I guess the problem is that libsvm is not properly linked, As i use -l with proper flags to compile with nvcc, but i don't know what to use with -l to properly link libsvm and use it.
i downloaded libsvm using
!git clone https://github.com/cjlin1/libsvm
%cd libsvm/
!make && make install
%cd /content/libsvm/python/
!make
import sys
sys.path.append('/content/libsvm/python')
%cd /content
now when i run this program
%%cuda --name Blind_Deblurring_Cuda.cu
#include <iostream>
#include <fstream>
#include <iostream>
#include <fstream>
#include "/content/brisque.h"
#include "/content/libsvm/svm.h"
#include <vector>
#include <stdio.h>
#include "fstream"
#include "iostream"
#include <algorithm>
#include <iterator>
#include <cmath>
#include<stdlib.h>
#include <math.h>
#include <curand.h>
#include <opencv2/core/cuda.hpp>
#include <opencv2/core.hpp>
#include "opencv2/imgproc.hpp"
#include "opencv2/imgcodecs.hpp"
#include <opencv2/core/core.hpp>
#include <iostream>
#include "opencv2/highgui.hpp"
#include <opencv2/core/utility.hpp>
//rescaling based on training data i libsvm
float rescale_vector[36][2];
using namespace std;
using namespace cv;
float computescore(string imagename);
void ComputeBrisqueFeature(Mat& orig, vector<double>& featurevector);
int read_range_file() {
//check if file exists
char buff[100];
int i;
string range_fname = "allrange";
FILE* range_file = fopen(range_fname.c_str(), "r");
if(range_file == NULL) return 1;
//assume standard file format for this program
fgets(buff, 100, range_file);
fgets(buff, 100, range_file);
//now we can fill the array
for(i = 0; i < 36; ++i) {
float a, b, c;
fscanf(range_file, "%f %f %f", &a, &b, &c);
rescale_vector[i][0] = b;
rescale_vector[i][1] = c;
}
return 0;
}
int main(int argc, char** argv)
{
if(argc < 2) {
cout << "Input Image argument not given." << endl;
return -1;
}
//read in the allrange file to setup internal scaling array
if(read_range_file()) {
cerr<<"unable to open allrange file"<<endl;
return -1;
}
float qualityscore;
qualityscore = computescore(argv[1]);
cout << "Quality Score: " << qualityscore << endl;
}
float computescore(string imagename) {
// pre-loaded vectors from allrange file
float min_[36] = {0.336999 ,0.019667 ,0.230000 ,-0.125959 ,0.000167 ,0.000616 ,0.231000 ,-0.125873 ,0.000165 ,0.000600 ,0.241000 ,-0.128814 ,0.000179 ,0.000386 ,0.243000 ,-0.133080 ,0.000182 ,0.000421 ,0.436998 ,0.016929 ,0.247000 ,-0.200231 ,0.000104 ,0.000834 ,0.257000 ,-0.200017 ,0.000112 ,0.000876 ,0.257000 ,-0.155072 ,0.000112 ,0.000356 ,0.258000 ,-0.154374 ,0.000117 ,0.000351};
float max_[36] = {9.999411, 0.807472, 1.644021, 0.202917, 0.712384, 0.468672, 1.644021, 0.169548, 0.713132, 0.467896, 1.553016, 0.101368, 0.687324, 0.533087, 1.554016, 0.101000, 0.689177, 0.533133, 3.639918, 0.800955, 1.096995, 0.175286, 0.755547, 0.399270, 1.095995, 0.155928, 0.751488, 0.402398, 1.041992, 0.093209, 0.623516, 0.532925, 1.042992, 0.093714, 0.621958, 0.534484};
double qualityscore;
int i;
struct svm_model* model; // create svm model object
Mat orig = imread(imagename, 1); // read image (color mode)
vector<double> brisqueFeatures; // feature vector initialization
ComputeBrisqueFeature(orig, brisqueFeatures); // compute brisque features
// use the pre-trained allmodel file
string modelfile = "allmodel";
//if((model=svm_load_model(modelfile.c_str()))==0) {
//fprintf(stderr,"can't open model file allmodel\n");
// exit(1);
//}
// float min_[37];
// float max_[37];
struct svm_node x[37];
// rescale the brisqueFeatures vector from -1 to 1
// also convert vector to svm node array object
for(i = 0; i < 36; ++i) {
float min = min_[i];
float max = max_[i];
x[i].value = -1 + (2.0/(max - min) * (brisqueFeatures[i] - min));
x[i].index = i + 1;
}
x[36].index = -1;
int nr_class=svm_get_nr_class(model);
double *prob_estimates = (double *) malloc(nr_class*sizeof(double));
// predict quality score using libsvm class
qualityscore = svm_predict_probability(model,x,prob_estimates);
free(prob_estimates);
svm_free_and_destroy_model(&model);
return qualityscore;
}
void ComputeBrisqueFeature(Mat& orig, vector<double>& featurevector)
{
Mat orig_bw_int(orig.size(), CV_64F, 1);
// convert to grayscale
cvtColor(orig, orig_bw_int, COLOR_BGR2GRAY);
// create a copy of original image
Mat orig_bw(orig_bw_int.size(), CV_64FC1, 1);
orig_bw_int.convertTo(orig_bw, 1.0/255);
orig_bw_int.release();
// orig_bw now contains the grayscale image normalized to the range 0,1
int scalenum = 2; // number of times to scale the image
for (int itr_scale = 1; itr_scale<=scalenum; itr_scale++)
{
// resize image
Size dst_size(orig_bw.cols/cv::pow((double)2, itr_scale-1), orig_bw.rows/pow((double)2, itr_scale-1));
Mat imdist_scaled;
resize(orig_bw, imdist_scaled, dst_size, 0, 0, INTER_CUBIC); // INTER_CUBIC
imdist_scaled.convertTo(imdist_scaled, CV_64FC1, 1.0/255.0);
// calculating MSCN coefficients
// compute mu (local mean)
Mat mu(imdist_scaled.size(), CV_64FC1, 1);
GaussianBlur(imdist_scaled, mu, Size(7, 7), 1.166);
Mat mu_sq;
cv::pow(mu, double(2.0), mu_sq);
//compute sigma (local sigma)
Mat sigma(imdist_scaled.size(), CV_64FC1, 1);
cv::multiply(imdist_scaled, imdist_scaled, sigma);
GaussianBlur(sigma, sigma, Size(7, 7), 1.166);
cv::subtract(sigma, mu_sq, sigma);
cv::pow(sigma, double(0.5), sigma);
add(sigma, Scalar(1.0/255), sigma); // to avoid DivideByZero Error
Mat structdis(imdist_scaled.size(), CV_64FC1, 1);
subtract(imdist_scaled, mu, structdis);
divide(structdis, sigma, structdis); // structdis is MSCN image
// Compute AGGD fit to MSCN image
double lsigma_best, rsigma_best, gamma_best;
structdis = AGGDfit(structdis, lsigma_best, rsigma_best, gamma_best);
featurevector.push_back(gamma_best);
featurevector.push_back((lsigma_best*lsigma_best + rsigma_best*rsigma_best)/2);
// Compute paired product images
// indices for orientations (H, V, D1, D2)
int shifts[4][2]={{0,1},{1,0},{1,1},{-1,1}};
for(int itr_shift=1; itr_shift<=4; itr_shift++)
{
// select the shifting index from the 2D array
int* reqshift = shifts[itr_shift-1];
// declare shifted_structdis as pairwise image
Mat shifted_structdis(imdist_scaled.size(), CV_64F, 1);
// create copies of the images using BwImage constructor
// utility constructor for better subscript access (for pixels)
BwImage OrigArr(structdis);
BwImage ShiftArr(shifted_structdis);
// create pair-wise product for the given orientation (reqshift)
for(int i=0; i<structdis.rows; i++)
{
for(int j=0; j<structdis.cols; j++)
{
if(i+reqshift[0]>=0 && i+reqshift[0]<structdis.rows && j+reqshift[1]>=0 && j+reqshift[1]<structdis.cols)
{
ShiftArr[i][j]=OrigArr[i + reqshift[0]][j + reqshift[1]];
}
else
{
ShiftArr[i][j]=0;
}
}
}
// Mat structdis_pairwise;
shifted_structdis = ShiftArr.equate(shifted_structdis);
// calculate the products of the pairs
multiply(structdis, shifted_structdis, shifted_structdis);
// fit the pairwise product to AGGD
shifted_structdis = AGGDfit(shifted_structdis, lsigma_best, rsigma_best, gamma_best);
double constant = sqrt(tgamma(1/gamma_best))/sqrt(tgamma(3/gamma_best));
double meanparam = (rsigma_best-lsigma_best)*(tgamma(2/gamma_best)/tgamma(1/gamma_best))*constant;
// push the calculated parameters from AGGD fit to pair-wise products
featurevector.push_back(gamma_best);
featurevector.push_back(meanparam);
featurevector.push_back(cv::pow(lsigma_best,2));
featurevector.push_back(cv::pow(rsigma_best,2));
}
}
}
// function to compute best fit parameters from AGGDfit
Mat AGGDfit(Mat structdis, double& lsigma_best, double& rsigma_best, double& gamma_best)
{
// create a copy of an image using BwImage constructor (brisque.h - more info)
BwImage ImArr(structdis);
long int poscount=0, negcount=0;
double possqsum=0, negsqsum=0, abssum=0;
for(int i=0;i<structdis.rows;i++)
{
for (int j =0; j<structdis.cols; j++)
{
double pt = ImArr[i][j]; // BwImage provides [][] access
if(pt>0)
{
poscount++;
possqsum += pt*pt;
abssum += pt;
}
else if(pt<0)
{
negcount++;
negsqsum += pt*pt;
abssum -= pt;
}
}
}
lsigma_best = cv::pow(negsqsum/negcount, 0.5);
rsigma_best = cv::pow(possqsum/poscount, 0.5);
double gammahat = lsigma_best/rsigma_best;
long int totalcount = (structdis.cols)*(structdis.rows);
double rhat = cv::pow(abssum/totalcount, static_cast<double>(2))/((negsqsum + possqsum)/totalcount);
double rhatnorm = rhat*(cv::pow(gammahat,3) +1)*(gammahat+1)/pow(pow(gammahat,2)+1,2);
double prevgamma = 0;
double prevdiff = 1e10;
float sampling = 0.001;
for (float gam=0.2; gam<10; gam+=sampling) //possible to coarsen sampling to quicken the code, with some loss of accuracy
{
double r_gam = tgamma(2/gam)*tgamma(2/gam)/(tgamma(1/gam)*tgamma(3/gam));
double diff = abs(r_gam-rhatnorm);
if(diff> prevdiff) break;
prevdiff = diff;
prevgamma = gam;
}
gamma_best = prevgamma;
return structdis.clone();
}
And then try to compile using
!nvcc -o /content/src/Blind_Deblurring_Cuda /content/src/Blind_Deblurring_Cuda.cu -lopencv_core -lopencv_imgcodecs -lopencv_imgproc -lopencv_highgui -lopencv_ml
It gives the following error
/tmp/tmpxft_00003d8d_00000000-10_Blind_Deblurring_Cuda.o: In function `computescore(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >)':
tmpxft_00003d8d_00000000-5_Blind_Deblurring_Cuda.cudafe1.cpp:(.text+0x9bc): undefined reference to `svm_get_nr_class'
tmpxft_00003d8d_00000000-5_Blind_Deblurring_Cuda.cudafe1.cpp:(.text+0x9fd): undefined reference to `svm_predict_probability'
tmpxft_00003d8d_00000000-5_Blind_Deblurring_Cuda.cudafe1.cpp:(.text+0xa27): undefined reference to `svm_free_and_destroy_model'
collect2: error: ld returned 1 exit status
I have a byte I am using to store bit flags. I need to compute the position of the most significant set bit in the byte.
Example Byte: 00101101 => 6 is the position of the most significant set bit
Compact Hex Mapping:
[0x00] => 0x00
[0x01] => 0x01
[0x02,0x03] => 0x02
[0x04,0x07] => 0x03
[0x08,0x0F] => 0x04
[0x10,0x1F] => 0x05
[0x20,0x3F] => 0x06
[0x40,0x7F] => 0x07
[0x80,0xFF] => 0x08
TestCase in C:
#include <stdio.h>
unsigned char check(unsigned char b) {
unsigned char c = 0x08;
unsigned char m = 0x80;
do {
if(m&b) { return c; }
else { c -= 0x01; }
} while(m>>=1);
return 0; //never reached
}
int main() {
unsigned char input[256] = {
0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f,
0x10,0x11,0x12,0x13,0x14,0x15,0x16,0x17,0x18,0x19,0x1a,0x1b,0x1c,0x1d,0x1e,0x1f,
0x20,0x21,0x22,0x23,0x24,0x25,0x26,0x27,0x28,0x29,0x2a,0x2b,0x2c,0x2d,0x2e,0x2f,
0x30,0x31,0x32,0x33,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x3b,0x3c,0x3d,0x3e,0x3f,
0x40,0x41,0x42,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x4b,0x4c,0x4d,0x4e,0x4f,
0x50,0x51,0x52,0x53,0x54,0x55,0x56,0x57,0x58,0x59,0x5a,0x5b,0x5c,0x5d,0x5e,0x5f,
0x60,0x61,0x62,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x6b,0x6c,0x6d,0x6e,0x6f,
0x70,0x71,0x72,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x7b,0x7c,0x7d,0x7e,0x7f,
0x80,0x81,0x82,0x83,0x84,0x85,0x86,0x87,0x88,0x89,0x8a,0x8b,0x8c,0x8d,0x8e,0x8f,
0x90,0x91,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0x9b,0x9c,0x9d,0x9e,0x9f,
0xa0,0xa1,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xab,0xac,0xad,0xae,0xaf,
0xb0,0xb1,0xb2,0xb3,0xb4,0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xbb,0xbc,0xbd,0xbe,0xbf,
0xc0,0xc1,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xcb,0xcc,0xcd,0xce,0xcf,
0xd0,0xd1,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xdb,0xdc,0xdd,0xde,0xdf,
0xe0,0xe1,0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xeb,0xec,0xed,0xee,0xef,
0xf0,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa,0xfb,0xfc,0xfd,0xfe,0xff };
unsigned char truth[256] = {
0x00,0x01,0x02,0x02,0x03,0x03,0x03,0x03,0x04,0x04,0x04,0x04,0x04,0x04,0x04,0x04,
0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,0x05,
0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,
0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,0x06,
0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,
0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,
0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,
0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,0x07,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,
0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08,0x08};
int i,r;
int f = 0;
for(i=0; i<256; ++i) {
r=check(input[i]);
if(r !=(truth[i])) {
printf("failed %d : 0x%x : %d\n",i,0x000000FF & ((int)input[i]),r);
f += 1;
}
}
if(!f) { printf("passed all\n"); }
else { printf("failed %d\n",f); }
return 0;
}
I would like to simplify my check() function to not involve looping (or branching preferably). Is there a bit twiddling hack or hashed lookup table solution to compute the position of the most significant set bit in a byte?
Your question is about an efficient way to compute log2 of a value. And because you seem to want a solution that is not limited to the C language I have been slightly lazy and tweaked some C# code I have.
You want to compute log2(x) + 1 and for x = 0 (where log2 is undefined) you define the result as 0 (e.g. you create a special case where log2(0) = -1).
static readonly Byte[] multiplyDeBruijnBitPosition = new Byte[] {
7, 2, 3, 4,
6, 1, 5, 0
};
public static Byte Log2Plus1(Byte value) {
if (value == 0)
return 0;
var roundedValue = value;
roundedValue |= (Byte) (roundedValue >> 1);
roundedValue |= (Byte) (roundedValue >> 2);
roundedValue |= (Byte) (roundedValue >> 4);
var log2 = multiplyDeBruijnBitPosition[((Byte) (roundedValue*0xE3)) >> 5];
return (Byte) (log2 + 1);
}
This bit twiddling hack is taken from Find the log base 2 of an N-bit integer in O(lg(N)) operations with multiply and lookup where you can see the equivalent C source code for 32 bit values. This code has been adapted to work on 8 bit values.
However, you may be able to use an operation that gives you the result using a very efficient built-in function (on many CPU's a single instruction like the Bit Scan Reverse is used). An answer to the question Bit twiddling: which bit is set? has some information about this. A quote from the answer provides one possible reason why there is low level support for solving this problem:
Things like this are the core of many O(1) algorithms such as kernel schedulers which need to find the first non-empty queue signified by an array of bits.
That was a fun little challenge. I don't know if this one is completely portable since I only have VC++ to test with, and I certainly can't say for sure if it's more efficient than other approaches. This version was coded with a loop but it can be unrolled without too much effort.
static unsigned char check(unsigned char b)
{
unsigned char r = 8;
unsigned char sub = 1;
unsigned char s = 7;
for (char i = 0; i < 8; i++)
{
sub = sub & ((( b & (1 << s)) >> s--) - 1);
r -= sub;
}
return r;
}
I'm sure everyone else has long since moved on to other topics but there was something in the back of my mind suggesting that there had to be a more efficient branch-less solution to this than just unrolling the loop in my other posted solution. A quick trip to my copy of Warren put me on the right track: Binary search.
Here's my solution based on that idea:
Pseudo-code:
// see if there's a bit set in the upper half
if ((b >> 4) != 0)
{
offset = 4;
b >>= 4;
}
else
offset = 0;
// see if there's a bit set in the upper half of what's left
if ((b & 0x0C) != 0)
{
offset += 2;
b >>= 2;
}
// see if there's a bit set in the upper half of what's left
if > ((b & 0x02) != 0)
{
offset++;
b >>= 1;
}
return b + offset;
Branch-less C++ implementation:
static unsigned char check(unsigned char b)
{
unsigned char adj = 4 & ((((unsigned char) - (b >> 4) >> 7) ^ 1) - 1);
unsigned char offset = adj;
b >>= adj;
adj = 2 & (((((unsigned char) - (b & 0x0C)) >> 7) ^ 1) - 1);
offset += adj;
b >>= adj;
adj = 1 & (((((unsigned char) - (b & 0x02)) >> 7) ^ 1) - 1);
return (b >> adj) + offset + adj;
}
Yes, I know that this is all academic :)
It is not possible in plain C. The best I would suggest is the following implementation of check. Despite quite "ugly" I think it runs faster than the ckeck version in the question.
int check(unsigned char b)
{
if(b&128) return 8;
if(b&64) return 7;
if(b&32) return 6;
if(b&16) return 5;
if(b&8) return 4;
if(b&4) return 3;
if(b&2) return 2;
if(b&1) return 1;
return 0;
}
Edit: I found a link to the actual code: http://www.hackersdelight.org/hdcodetxt/nlz.c.txt
The algorithm below is named nlz8 in that file. You can choose your favorite hack.
/*
From last comment of: http://stackoverflow.com/a/671826/315052
> Hacker's Delight explains how to correct for the error in 32-bit floats
> in 5-3 Counting Leading 0's. Here's their code, which uses an anonymous
> union to overlap asFloat and asInt: k = k & ~(k >> 1); asFloat =
> (float)k + 0.5f; n = 158 - (asInt >> 23); (and yes, this relies on
> implementation-defined behavior) - Derrick Coetzee Jan 3 '12 at 8:35
*/
unsigned char check (unsigned char b) {
union {
float asFloat;
int asInt;
} u;
unsigned k = b & ~(b >> 1);
u.asFloat = (float)k + 0.5f;
return 32 - (158 - (u.asInt >> 23));
}
Edit -- not exactly sure what the asker means by language independent, but below is the equivalent code in python.
import ctypes
class Anon(ctypes.Union):
_fields_ = [
("asFloat", ctypes.c_float),
("asInt", ctypes.c_int)
]
def check(b):
k = int(b) & ~(int(b) >> 1)
a = Anon(asFloat=(float(k) + float(0.5)))
return 32 - (158 - (a.asInt >> 23))