How to optimize my c code? - c

Question

I tried to implement C code for Wavelet transform in FPGA (Zynq ZC 702) but the code get stuck and this is because of memory problem so I should optimize my code but I don't know how.
Can anyone please give me some ideas how to do that ?
This is the main of the code
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include "wavemin.h"
#include "waveaux.h"
#include "waveaux.c"
#include "wavemin.c"
int main() {
printf("Hello World1 \n\r");
wave_object obj;
wt_object wt;
float *inp, *out;
int N, i, J,k;
float temp[1280] = {};
char *name = "db4";
obj = wave_init(name);
printf("Hello World2 \n\r");
N = 1280;
inp = (float*)malloc(sizeof(float) * N);
out = (float*)malloc(sizeof(float) * N);
//wmean = mean(temp, N);
for (i = 0; i < N; ++i) {
inp[i] = temp[i];
printf("Hello World3 \n\r");
//printf("%g \n", inp[i]);
}
J = 4; //Decomposition Levels
wt = wt_init(obj, "dwt", N, J); // Initialize the wavelet transform object
printf("Hello World4 \n\r");
setDWTExtension(wt, "sym"); // Options are "per" and "sym". Symmetric is the default option
printf("Hello World5 \n\r");
setWTConv(wt, "direct");
printf("Hello World6 \n\r");
dwt(wt, inp); // Perform DWT
printf("Hello World7 \n\r");
//getDWTAppx(wt, out, wt->length[0]);
// printf("Approximation Coefficients Level 1 \n");
// for (i = 0; i < wt->length[0]; ++i) {
// printf("%g ", out[i]);
// }
// printf("\n\n");
for (k = 1; k <= J; ++k) {
getDWTDetail(wt, out, wt->length[k], k);
printf("Detail Coefficients Level %d Length %d \n",
k, wt - length[k]);
for (i = 0; i < wt->length[k]; ++i) {
printf("%g ", out[i]);
}
printf("\n\n");
}
wt_summary(wt);// Prints the full summary.
printf("Hello World8 \n\r");
wave_free(obj);
wt_free(wt);
free(inp);
free(out);
return 0;
}
The other part of the code where there is the function used in the main function:
#include "wavemin.h"
wave_object wave_init(char *wname) {
wave_object obj = NULL;
int retval;
retval = 0;
if (wname != NULL) {
retval = filtlength(wname);
}
obj = (wave_object)malloc(sizeof(struct wave_set) + sizeof(float) * 4 *
retval);
obj->filtlength = retval;
obj->lpd_len = obj->hpd_len = obj->lpr_len = obj->hpr_len = obj->filtlength;
strcpy(obj->wname, wname);
if (wname != NULL) {
filtcoef(wname, obj->params, obj->params + retval, obj->params + 2 *
retval, obj->params + 3 * retval);
}
obj->lpd = &obj->params[0];
obj->hpd = &obj->params[retval];
obj->lpr = &obj->params[2 * retval];
obj->hpr = &obj->params[3 * retval];
return obj;
}
wt_object wt_init(wave_object wave, char *method, int siglength, int J) {
int size, i, MaxIter;
wt_object obj = NULL;
size = wave->filtlength;
MaxIter = wmaxiter(siglength, size);
if (!strcmp(method, "dwt") || !strcmp(method, "DWT")) {
obj = (wt_object)malloc(sizeof(struct wt_set) + sizeof(float) *
(siglength + 2 * J * (size + 1)));
obj->outlength = siglength + 2 * J * (size + 1); // Default
strcpy(obj->ext, "sym"); // Default
}
obj->wave = wave;
obj->siglength = siglength;
obj->J = J;
obj->MaxIter = MaxIter;
strcpy(obj->method, method);
if (siglength % 2 == 0) {
obj->even = 1;
}
else {
obj->even = 0;
}
strcpy(obj->cmethod, "direct"); // Default
obj->cfftset = 0;
obj->lenlength = J + 2;
obj->output = &obj->params[0];
if (!strcmp(method, "dwt") || !strcmp(method, "DWT")) {
for (i = 0; i < siglength + 2 * J * (size + 1); ++i) {
obj->params[i] = 0.0;
}
}
//wave_summary(obj->wave);
return obj;
}
static void dwt_sym(wt_object wt, float *inp, int N, float *cA, int len_cA,
float *cD, int len_cD) {
int i, l, t, len_avg;
len_avg = wt->wave->lpd_len;
for (i = 0; i < len_cA; ++i) {
t = 2 * i + 1;
cA[i] = 0.0;
cD[i] = 0.0;
for (l = 0; l < len_avg; ++l) {
if ((t - l) >= 0 && (t - l) < N) {
cA[i] += wt->wave->lpd[l] * inp[t - l];
cD[i] += wt->wave->hpd[l] * inp[t - l];
printf("world1 \n\r");
}
else if ((t - l) < 0) {
cA[i] += wt->wave->lpd[l] * inp[-t + l - 1];
cD[i] += wt->wave->hpd[l] * inp[-t + l - 1];
printf("world2 \n\r");
}
else if ((t - l) >= N) {
cA[i] += wt->wave->lpd[l] * inp[2 * N - t + l - 1];
cD[i] += wt->wave->hpd[l] * inp[2 * N - t + l - 1];
printf("world3 \n\r");
}
}
}
}
void dwt(wt_object wt, float *inp) {
int i, J, temp_len, iter, N, lp;
int len_cA;
float *orig, *orig2;
temp_len = wt->siglength;
J = wt->J;
wt->length[J + 1] = temp_len;
wt->outlength = 0;
wt->zpad = 0;
orig = (float*)malloc(sizeof(float) * temp_len);
orig2 = (float*)malloc(sizeof(float) * temp_len);
for (i = 0; i < wt->siglength; ++i) {
orig[i] = inp[i];
printf("Hello1 \n\r");
}
if (wt->zpad == 1) {
orig[temp_len - 1] = orig[temp_len - 2];
printf("Hello2 \n\r");
}
N = temp_len;
lp = wt->wave->lpd_len;
if (!strcmp(wt->ext, "sym")) {
//printf("\n YES %s \n", wt->ext);
i = J;
while (i > 0) {
N = N + lp - 2;
N = (int)ceil((float)N / 2.0);
wt->length[i] = N;
wt->outlength += wt->length[i];
i--;
}
wt->length[0] = wt->length[1];
wt->outlength += wt->length[0];
N = wt->outlength;
printf("Hello3 \n\r");
for (iter = 0; iter < J; ++iter) {
len_cA = wt->length[J - iter];
N -= len_cA;
dwt_sym(wt, orig, temp_len, orig2, len_cA, wt->params + N, len_cA);
temp_len = wt->length[J - iter];
printf("Hello4 \n\r");
if (iter == J - 1) {
for (i = 0; i < len_cA; ++i) {
wt->params[i] = orig2[i];
printf("Hello5 \n\r");
}
} else {
for (i = 0; i < len_cA; ++i) {
orig[i] = orig2[i];
printf("Hello6 \n\r");
}
}
}
} else {
printf("Signal extension can be either per or sym");
exit(-1);
}
free(orig);
free(orig2);
}
void setDWTExtension(wt_object wt, char *extension) {
if (!strcmp(extension, "sym")) {
strcpy(wt->ext, "sym");
} else {
printf("Signal extension can be either per or sym");
exit(-1);
}
}
void setWTConv(wt_object wt, char *cmethod) {
if (!strcmp(cmethod, "direct")) {
strcpy(wt->cmethod, "direct");
}
}
void getDWTDetail(wt_object wt, float *detail, int N, int level) {
/*
returns Detail coefficents at the jth level where j = 1,2,.., J
and Wavelet decomposition is stored as
[A(J) D(J) D(J-1) ..... D(1)] in wt->output vector
Use getDWTAppx() to get A(J)
Level 1 : Length of D(J), ie N, is stored in wt->length[1]
Level 2 :Length of D(J-1), ie N, is stored in wt->length[2]
....
Level J : Length of D(1), ie N, is stored in wt->length[J]
*/
int i, iter, J;
J = wt->J;
if (level > J) {
printf("The decomposition only has %d levels", J);
}
iter = wt->length[0];
for (i = 1; i < level; ++i) {
iter += wt->length[i];
}
for (i = 0; i < N; ++i) {
detail[i] = wt->output[i + iter];
}
}
void getDWTAppx(wt_object wt, float *appx, int N) {
/*
Wavelet decomposition is stored as
[A(J) D(J) D(J-1) ..... D(1)] in wt->output vector
Length of A(J) , N = wt->length[0]
*/
int i;
for (i = 0; i < N; ++i) {
appx[i] = wt->output[i];
}
}
void wt_summary(wt_object wt) {
int i;
int J, t;
J = wt->J;
printf("Wavelet Coefficients are contained in vector : %s \n", "output");
printf("\n");
printf("Approximation Coefficients \n");
printf("Level %d Access : output[%d] Length : %d \n",
1, 0, wt->length[0]);
printf("\n");
printf("Detail Coefficients \n");
t = wt->length[0];
for (i = 0; i < J; ++i) {
printf("Level %d Access : output[%d] Length : %d \n",
i + 1, t, wt->length[i + 1]);
t += wt->length[i + 1];
}
printf("\n");
}
void wave_free(wave_object object) {
free(object);
}
void wt_free(wt_object object) {
free(object);
}
enter image description here

Answer 1

In your code
Always check if malloc has returned non NULL value
Check your stack and heap settings in the linker file as you declare massive local variables and do a lots of mallocs - I suspect the (nomen omen)stack overflow, or failed mallocs.
Is it a bare metal program or you run it under some kind of OS?

Answer 2

Just for a matter of style and concision, I would rewrite this:
if (siglength % 2 == 0) {
obj->even = 1;
}
else {
obj->even = 0;
}
Into the following code:
obj->even = !(siglength % 2);
Or, alternatively:
obj->even = (siglength % 2) ? 0 : 1;
Also, I think there is room for optimization in this function:
static void dwt_sym(wt_object wt, float *inp, int N, float *cA, int len_cA,
float *cD, int len_cD) {
int i, l, t, len_avg;
len_avg = wt->wave->lpd_len;
for (i = 0; i < len_cA; ++i) {
t = 2 * i + 1;
cA[i] = 0.0;
cD[i] = 0.0;
for (l = 0; l < len_avg; ++l) {
if ((t - l) >= 0 && (t - l) < N) {
cA[i] += wt->wave->lpd[l] * inp[t - l];
cD[i] += wt->wave->hpd[l] * inp[t - l];
printf("world1 \n\r");
}
else if ((t - l) < 0) {
cA[i] += wt->wave->lpd[l] * inp[-t + l - 1];
cD[i] += wt->wave->hpd[l] * inp[-t + l - 1];
printf("world2 \n\r");
}
else if ((t - l) >= N) {
cA[i] += wt->wave->lpd[l] * inp[2 * N - t + l - 1];
cD[i] += wt->wave->hpd[l] * inp[2 * N - t + l - 1];
printf("world3 \n\r");
}
}
}
}
First, you are always referring to t - 1 and never t itself, so why not have:
t = 2 * i;
And, I can guess that a lot of computation can be placed outside of the inner loop... If you want to optimize, there are many good candidate here.
One last word about optimization!
You should first profile your software and see where you spend the most time before thinking about optimization. You cannot optimize "in the air" without knowing where your software does really struggle. Consider using gprof.
PS: You should never ever use the letter l (ell) as a variable... it is way to close from the number 1 (one). Consider changing this is also, it can improve the reading.