I have set a function to setup the accelerator, after i have read :
Using the Apple FFT and Accelerate Framework
iPhone FFT with Accelerate framework vDSP
and apple docs.
i did this :
void fftSetup()
{
COMPLEX_SPLIT A;
FFTSetup setupReal;
uint32_t log2n;
uint32_t n, nOver2;
int32_t stride;
uint32_t i;
float *originalReal, *obtainedReal;
float scale;
uint32_t L = 1024;
float *mag = new float[L/2];
log2n = 10 ;
n = 1 << log2n;
stride = 1;
nOver2 = n / 2;
printf("1D real FFT of length log2 ( %d ) = %d\n\n", n, log2n);
for (i = 0; i < n; i++)
originalReal[i] = (float) (i + 1);
vDSP_ctoz((COMPLEX *) originalReal,2,&A,1,nOver2);
A.realp = (float *) malloc(nOver2 * sizeof(float));
A.imagp = (float *) malloc(nOver2 * sizeof(float));
setupReal = vDSP_create_fftsetup(log2n, FFT_RADIX2);
vDSP_fft_zrip(setupReal, &A, stride, log2n, FFT_FORWARD);
vDSP_fft_zrip(setupReal, &A, stride, log2n, FFT_INVERSE);
//get magnitude;
for(i = 1; i < L/2; i++){
mag[i] = sqrtf(A.realp[i]*A.realp[i] + A.imagp[i] * A.imagp[i]);
}
scale = (float) 1.0 / (2 * n);
vDSP_vsmul(A.realp, 1, &scale, A.realp, 1, nOver2);
vDSP_vsmul(A.imagp, 1, &scale, A.imagp, 1, nOver2);
}
questions :
my app is always crash with no error(BAD ACCESS) on one of this 2 lines :
originalReal[i] = (float) (i + 1); // or
vDSP_ctoz((COMPLEX *) originalReal,2,&A,1,nOver2);
i guess i did not set a good value for log2n ? (10 to get 1024 window ? )
how do i get the real magnitude of the bins? my actual fft? the same i wrote here ?
where do i input MY data buffer array (exactly where in my code ? instead originalReal?)
thanks a lot.
I actually manage to make it work ,when i insert into it a sin wave of a certain f.
This is the code :
COMPLEX_SPLIT A;
FFTSetup setupReal;
uint32_t log2n;
uint32_t n, nOver2;
int32_t stride;
uint32_t i;
float *originalReal, *obtainedReal;
float scale;
uint32_t L = 1024;
float *mag = new float[L/2];
log2n = 10 ;
n = 1 << log2n;
stride = 1;
nOver2 = n / 2;
//printf("1D real FFT of length log2 ( %d ) = %d\n\n", n, log2n);
A.realp = (float *) malloc(nOver2 * sizeof(float));
A.imagp = (float *) malloc(nOver2 * sizeof(float));
originalReal = (float *) malloc(n * sizeof(float));
obtainedReal = (float *) malloc(n * sizeof(float));
for (i = 0; i < n; i++)
originalReal[i] = cos(2*3.141592*11000*i/44100);//(float) (i + 1);
vDSP_ctoz((COMPLEX *) originalReal,2,&A,1,nOver2);
setupReal = vDSP_create_fftsetup(log2n, FFT_RADIX2);
vDSP_fft_zrip(setupReal, &A, stride, log2n, FFT_FORWARD);
//vDSP_fft_zrip(setupReal, &A, stride, log2n, FFT_INVERSE);
scale = (float) 1.0 / (2 * n);
vDSP_vsmul(A.realp, 1, &scale, A.realp, 1, nOver2);
vDSP_vsmul(A.imagp, 1, &scale, A.imagp, 1, nOver2);
//get magnitude;
for(i = 1; i < L/2; i++)
{
mag[i] = sqrtf(A.realp[i]*A.realp[i] + A.imagp[i] * A.imagp[i]);
NSLog(#"%d:%f",i,mag[i]);
}
Actually its not 44hz between bins,as the guy wrote in the post above! but 43 ! 22050/512=43 . this thing is critical ! because in the higher bins- such as bin[300] you get a completely different resault for 44 and 43 ! (its 300hz drift). so take care of that .
Related
I am a beginner in doing GPU programming with OpenACC. I was trying to do a direct convolution. Convolution consists of 6 nested loops. I only want the first loop to be parallelized. I gave the pragma #pragma acc loop for the first loop and #pragma acc loop seq for the rest. But the output that I am getting is not correct. Is the approach taken by me to parallelize the loop correct ? Specifications for the convolution: Input channels-3, Input Size- 224X224X3, Output channels- 64, Output Size- 111X111X64, filter size- 3X3X3X64. Following is the link to the header files dog.h and squeezenet_params.h. https://drive.google.com/drive/folders/1a9XRjBTrEFIorrLTPFHS4atBOPrG886i
# include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "squeezenet_params.h"
#include "dog.h"
void conv3x3(
const int input_channels, const int input_size,
const int pad, const int stride, const int start_channel,
const int output_size, const float* restrict input_im, const float* restrict filter_weight,
const float* restrict filter_bias, float* restrict output_im){
#pragma acc data copyin (input_im[0:150527],filter_weight[0:1727],filter_bias[0:63]) copyout(output_im[0:788543])
{
#pragma acc parallel
{
#pragma acc loop
for(int p=0;p<64;++p){
filter_weight += p * input_channels * 9;
float bias = filter_bias[p];
output_im += (start_channel + p) * output_size * output_size;
//loop over output feature map
#pragma acc loop seq
for(int i = 0; i < output_size; i++)
{
#pragma acc loop seq
for(int j = 0; j < output_size; j++)
{
//compute one element in the output feature map
float tmp = bias;
//compute dot product of 2 input_channels x 3 x 3 matrix
#pragma acc loop seq
for(int k = 0; k < input_channels; k++)
{
#pragma acc loop seq
for(int l = 0; l < 3; l++)
{
int h = i * stride + l - pad;
#pragma acc loop seq
for(int m = 0; m < 3; m++)
{
int w = j * stride + m - pad;
if((h >= 0) && (h < input_size) && (w >= 0) && (w < input_size))
{
tmp += input_im[k * input_size * input_size + (i * stride + l - pad) * input_size + j * stride + m - pad] \
* filter_weight[9 * k + 3 * l + m];
}
}
}
}
//add relu activation after conv
output_im[i * output_size + j] = (tmp > 0.0) ? tmp : 0.0;
}
}
}
}
}
}
void main(){
float * result = (float*)malloc(sizeof(float) * (1 * 64 * 111 * 111));
conv3x3(3,224,0,2,0,111,sample,conv1_weight,conv1_bias,result);
for(int i=0;i<64 * 111 * 111;++i){
//if(result[i]>0)
printf("%f:%d\n",result[i],i);
}
}
The contributor posted the same question on the PGI User Forums where I've answered. (See: https://www.pgroup.com/userforum/viewtopic.php?f=4&t=7614). The topic question is incorrect in that the inner loops are not getting parallelized nor are the cause of the issue.
The problem here is that the code has a race condition on the shared "output_im" pointer. My suggested solution is to compute a per thread offset into the array rather than trying to manipulate the pointer itself.
for(int p=0;p<64;++p){
filter_weight += p * input_channels * 9;
float bias = filter_bias[p];
int offset;
offset = (start_channel + p) * output_size * output_size;
//loop over output feature map
#pragma acc loop vector collapse(2)
for(int i = 0; i < output_size; i++)
{
for(int j = 0; j < output_size; j++)
{
... cut ...
}
}
//add relu activation after conv
int idx = offset + (i * output_size + j);
output_im[idx] = (tmp > 0.0) ? tmp : 0.0;
}
}
I know this question is very noob. I am trying to understand how the pointer thing works. I studied basics of C but still did not understand this.
Given this piece of function:
+ (void)nv21ToRgbWithWidth:(unsigned int)width height:(unsigned int)height yuyv:(unsigned char *)yuyv rgb:(unsigned char *)rgb
{
const int nv_start = width * height ;
UInt32 i, j, index = 0, rgb_index = 0;
UInt8 y, u, v;
int r, g, b, nv_index = 0;
for(i = 0; i < height ; i++)
{
for(j = 0; j < width; j ++){
//nv_index = (rgb_index / 2 - width / 2 * ((i + 1) / 2)) * 2;
nv_index = i / 2 * width + j - j % 2;
y = yuyv[rgb_index];
u = yuyv[nv_start + nv_index ];
v = yuyv[nv_start + nv_index + 1];
r = y + (140 * (v-128))/100; //r
g = y - (34 * (u-128))/100 - (71 * (v-128))/100; //g
b = y + (177 * (u-128))/100; //b
if(r > 255) r = 255;
if(g > 255) g = 255;
if(b > 255) b = 255;
if(r < 0) r = 0;
if(g < 0) g = 0;
if(b < 0) b = 0;
index = rgb_index % width + (height - i - 1) * width;
rgb[index * 3+0] = b;
rgb[index * 3+1] = g;
rgb[index * 3+2] = r;
rgb_index++;
}
}
}
How am I suppose to know how the unsigned char * for rgb should be initialized before passing in to the function?
I tried calling the function like this:
unsigned char *rgb = NULL;
[MyClass nv21ToRgbWithWidth:imageWidth height:imageHeight yuyv:yuyvValues rgb:rgb];
But the the program crashes on this line:
rgb[index * 3+0] = b;
I see rgb was initialized with NULL, so you can't assign values. So, I thought of initializing an array and pass it to pointer rgb like this:
unsigned char rgbArr[10000];
unsigned char *rgb = rgbArr;
but the function still crashes. I really don't know how should I pass the rgb parameter in this function. Please help me understand this.
The expected size in bytes seems to be at least height*width*3; it might be that allocating such an array as a local variable (as you do with unsigned char rgbArr[10000]) exceeds a stack limit; The program likely crashes in such a case. I'd try to use the heap instead:
unsigned char* rgb = malloc(imageHeight*imageWidth*3);
[MyClass nv21ToRgbWithWidth:imageWidth height:imageHeight yuyv:yuyvValues rgb:rgb];
...
free(rgb);
That is what the malloc(), calloc(), realloc() and free() functions are for. Don't forget to use the free() function to prevent memory leaks... I hope that helps.
I'm trying to compute batch 1D FFTs using cufftPlanMany. The data set comes from a 3D field, stored in a 1D array, where I want to compute 1D FFTs in the x and y direction. The data is stored as shown in the figure below; continuous in x then y then z.
Doing batch FFTs in the x-direction is (I believe) straighforward; with input stride=1, distance=nx and batch=ny * nz, it computes the FFTs over elements {0,1,2,3}, {4,5,6,7}, ..., {28,29,30,31}. However, I can't think of a way to achieve the same for the FFTs in the y-direction. A batch for each xy plane is again straightforward (input stride=nx, dist=1, batch=nx results in FFTs over {0,4,8,12}, {1,5,9,13}, etc.). But with batch=nx * nz, going from {3,7,11,15} to {16,20,24,28}, the distance is larger than 1. Can this somehow be done with cufftPlanMany?
I think that the short answer to your question (possibility of using a single cufftPlanMany to perform 1D FFTs of the columns of a 3D matrix) is NO.
Indeed, transformations performed according to cufftPlanMany, that you call like
cufftPlanMany(&handle, rank, n,
inembed, istride, idist,
onembed, ostride, odist, CUFFT_C2C, batch);
must obey the Advanced Data Layout. In particular, 1D FFTs are worked out according to the following layout
input[b * idist + x * istride]
where b addresses the b-th signal and istride is the distance between two consecutive items in the same signal. If the 3D matrix has dimensions M * N * Q and if you want to perform 1D transforms along the columns, then the distance between two consecutive elements will be M, while the distance between two consecutive signals will be 1. Furthermore, the number of batched executions must be set equal to M. With those parameters, you are able to cover only one slice of the 3D matrix. Indeed, if you try increasing M, then the cuFFT will start trying to compute new column-wise FFTs starting from the second row. The only solution to this problem is an iterative call to cufftExecC2C to cover all the Q slices.
For the record, the following code provides a fully worked example on how performing 1D FFTs of the columns of a 3D matrix.
#include <thrust/device_vector.h>
#include <cufft.h>
/********************/
/* CUDA ERROR CHECK */
/********************/
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
int main() {
const int M = 3;
const int N = 4;
const int Q = 2;
thrust::host_vector<float2> h_matrix(M * N * Q);
for (int k=0; k<Q; k++)
for (int j=0; j<N; j++)
for (int i=0; i<M; i++) {
float2 temp;
temp.x = (float)(j + k * M);
//temp.x = 1.f;
temp.y = 0.f;
h_matrix[k*M*N+j*M+i] = temp;
printf("%i %i %i %f %f\n", i, j, k, temp.x, temp.y);
}
printf("\n");
thrust::device_vector<float2> d_matrix(h_matrix);
thrust::device_vector<float2> d_matrix_out(M * N * Q);
// --- Advanced data layout
// input[b * idist + x * istride]
// output[b * odist + x * ostride]
// b = signal number
// x = element of the b-th signal
cufftHandle handle;
int rank = 1; // --- 1D FFTs
int n[] = { N }; // --- Size of the Fourier transform
int istride = M, ostride = M; // --- Distance between two successive input/output elements
int idist = 1, odist = 1; // --- Distance between batches
int inembed[] = { 0 }; // --- Input size with pitch (ignored for 1D transforms)
int onembed[] = { 0 }; // --- Output size with pitch (ignored for 1D transforms)
int batch = M; // --- Number of batched executions
cufftPlanMany(&handle, rank, n,
inembed, istride, idist,
onembed, ostride, odist, CUFFT_C2C, batch);
for (int k=0; k<Q; k++)
cufftExecC2C(handle, (cufftComplex*)(thrust::raw_pointer_cast(d_matrix.data()) + k * M * N), (cufftComplex*)(thrust::raw_pointer_cast(d_matrix_out.data()) + k * M * N), CUFFT_FORWARD);
cufftDestroy(handle);
for (int k=0; k<Q; k++)
for (int j=0; j<N; j++)
for (int i=0; i<M; i++) {
float2 temp = d_matrix_out[k*M*N+j*M+i];
printf("%i %i %i %f %f\n", i, j, k, temp.x, temp.y);
}
}
The situation is different for the case when you want to perform 1D transforms of the rows. In that case, the distance between two consecutive elements is 1, while the distance between two consecutive signals is M. This allows you to set a number of N * Q transformations and then invoking cufftExecC2C only one time. For the record, the code below provides a full example of 1D transformations of the rows of a 3D matrix.
#include <thrust/device_vector.h>
#include <cufft.h>
/********************/
/* CUDA ERROR CHECK */
/********************/
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
int main() {
const int M = 3;
const int N = 4;
const int Q = 2;
thrust::host_vector<float2> h_matrix(M * N * Q);
for (int k=0; k<Q; k++)
for (int j=0; j<N; j++)
for (int i=0; i<M; i++) {
float2 temp;
temp.x = (float)(j + k * M);
//temp.x = 1.f;
temp.y = 0.f;
h_matrix[k*M*N+j*M+i] = temp;
printf("%i %i %i %f %f\n", i, j, k, temp.x, temp.y);
}
printf("\n");
thrust::device_vector<float2> d_matrix(h_matrix);
thrust::device_vector<float2> d_matrix_out(M * N * Q);
// --- Advanced data layout
// input[b * idist + x * istride]
// output[b * odist + x * ostride]
// b = signal number
// x = element of the b-th signal
cufftHandle handle;
int rank = 1; // --- 1D FFTs
int n[] = { M }; // --- Size of the Fourier transform
int istride = 1, ostride = 1; // --- Distance between two successive input/output elements
int idist = M, odist = M; // --- Distance between batches
int inembed[] = { 0 }; // --- Input size with pitch (ignored for 1D transforms)
int onembed[] = { 0 }; // --- Output size with pitch (ignored for 1D transforms)
int batch = N * Q; // --- Number of batched executions
cufftPlanMany(&handle, rank, n,
inembed, istride, idist,
onembed, ostride, odist, CUFFT_C2C, batch);
cufftExecC2C(handle, (cufftComplex*)(thrust::raw_pointer_cast(d_matrix.data())), (cufftComplex*)(thrust::raw_pointer_cast(d_matrix_out.data())), CUFFT_FORWARD);
cufftDestroy(handle);
for (int k=0; k<Q; k++)
for (int j=0; j<N; j++)
for (int i=0; i<M; i++) {
float2 temp = d_matrix_out[k*M*N+j*M+i];
printf("%i %i %i %f %f\n", i, j, k, temp.x, temp.y);
}
}
I guess, idist=nx*nz could also jump a whole plane and batch=nz would then cover one yx plane. The decision should be made according to whether nx or nz is larger.
Important update: I already figured out the answers and put them in this simple open-source library: http://bartolsthoorn.github.com/NVDSP/ Check it out, it will probably save you quite some time if you're having trouble with audio filters in IOS!
^
I have created a (realtime) audio buffer (float *data) that holds a few sin(theta) waves with different frequencies.
The code below shows how I created my buffer, and I've tried to do a bandpass filter but it just turns the signals to noise/blips:
// Multiple signal generator
__block float *phases = nil;
[audioManager setOutputBlock:^(float *data, UInt32 numFrames, UInt32 numChannels)
{
float samplingRate = audioManager.samplingRate;
NSUInteger activeSignalCount = [tones count];
// Initialize phases
if (phases == nil) {
phases = new float[10];
for(int z = 0; z <= 10; z++) {
phases[z] = 0.0;
}
}
// Multiple signals
NSEnumerator * enumerator = [tones objectEnumerator];
id frequency;
UInt32 c = 0;
while(frequency = [enumerator nextObject])
{
for (int i=0; i < numFrames; ++i)
{
for (int iChannel = 0; iChannel < numChannels; ++iChannel)
{
float theta = phases[c] * M_PI * 2;
if (c == 0) {
data[i*numChannels + iChannel] = sin(theta);
} else {
data[i*numChannels + iChannel] = data[i*numChannels + iChannel] + sin(theta);
}
}
phases[c] += 1.0 / (samplingRate / [frequency floatValue]);
if (phases[c] > 1.0) phases[c] = -1;
}
c++;
}
// Normalize data with active signal count
float signalMulti = 1.0 / (float(activeSignalCount) * (sqrt(2.0)));
vDSP_vsmul(data, 1, &signalMulti, data, 1, numFrames*numChannels);
// Apply master volume
float volume = masterVolumeSlider.value;
vDSP_vsmul(data, 1, &volume, data, 1, numFrames*numChannels);
if (fxSwitch.isOn) {
// H(s) = (s/Q) / (s^2 + s/Q + 1)
// http://www.musicdsp.org/files/Audio-EQ-Cookbook.txt
// BW 2.0 Q 0.667
// http://www.rane.com/note170.html
//The order of the coefficients are, B1, B2, A1, A2, B0.
float Fs = samplingRate;
float omega = 2*M_PI*Fs; // w0 = 2*pi*f0/Fs
float Q = 0.50f;
float alpha = sin(omega)/(2*Q); // sin(w0)/(2*Q)
// Through H
for (int i=0; i < numFrames; ++i)
{
for (int iChannel = 0; iChannel < numChannels; ++iChannel)
{
data[i*numChannels + iChannel] = (data[i*numChannels + iChannel]/Q) / (pow(data[i*numChannels + iChannel],2) + data[i*numChannels + iChannel]/Q + 1);
}
}
float b0 = alpha;
float b1 = 0;
float b2 = -alpha;
float a0 = 1 + alpha;
float a1 = -2*cos(omega);
float a2 = 1 - alpha;
float *coefficients = (float *) calloc(5, sizeof(float));
coefficients[0] = b1;
coefficients[1] = b2;
coefficients[2] = a1;
coefficients[3] = a2;
coefficients[3] = b0;
vDSP_deq22(data, 2, coefficients, data, 2, numFrames);
free(coefficients);
}
// Measure dB
[self measureDB:data:numFrames:numChannels];
}];
My aim is to make a 10-band EQ for this buffer, using vDSP_deq22, the syntax of the method is:
vDSP_deq22(<float *vDSP_A>, <vDSP_Stride vDSP_I>, <float *vDSP_B>, <float *vDSP_C>, <vDSP_Stride vDSP_K>, <vDSP_Length __vDSP_N>)
See: http://developer.apple.com/library/mac/#documentation/Accelerate/Reference/vDSPRef/Reference/reference.html#//apple_ref/doc/c_ref/vDSP_deq22
Arguments:
float *vDSP_A is the input data
float *vDSP_B are 5 filter coefficients
float *vDSP_C is the output data
I have to make 10 filters (10 times vDSP_deq22). Then I set the gain for every band and combine them back together. But what coefficients do I feed every filter? I know vDSP_deq22 is a 2nd order (butterworth) IIR filter, but how do I turn this into a bandpass?
Now I have three questions:
a) Do I have to de-interleave and interleave the audio buffer? I know setting stride to 2 just filters on channel but how I filter the other, stride 1 will process both channels as one.
b) Do I have to transform/process the buffer before it enters the vDSP_deq22 method? If so, do I also have to transform it back to normal?
c) What values of the coefficients should I set to the 10 vDSP_deq22s?
I've been trying for days now but I haven't been able to figure this on out, please help me out!
Your omega value need to be normalised, i.e. expressed as a fraction of Fs - it looks like you left out the f0 when you calculated omega, which will make alpha wrong too:
float omega = 2*M_PI*Fs; // w0 = 2*pi*f0/Fs
should probably be:
float omega = 2*M_PI*f0/Fs; // w0 = 2*pi*f0/Fs
where f0 is the centre frequency in Hz.
For your 10 band equaliser you'll need to pick 10 values of f0, spaced logarithmically, e.g. 25 Hz, 50 Hz, 100 Hz, 200 Hz, 400 Hz, 800 Hz, 1.6 kHz, 3.2 kHz, 6.4 kHz, 12.8 kHz.
I want to calculate the product A^T*A ( A is 2000x1000 Matrix). Also i only want to solve the upper triangular Matrix. In the inner loop i have to solve the dot product of two vectors.
Now, here is the problem. Using cblas ddot() is not faster than calculating the dot product with a loop. How is this possible? (using Intel Core (TM)i7 CPU M620 #2,67GHz, 1,92GB RAM)
The problem is caused essentially by matrix size, not by ddot. Your matrices are so large that they do not fit in the cache memory. The solution is to rearrange the three nested loops such that as much as possible can be done with a line in cache, so reducing cache refreshes. A model implementation follows for both the ddot and an daxpy approach. On my computer the time consumption was about 15:1.
In other words: never, never, never program a matrix multiplication along the "row times column" scheme that we learned in school.
/*
Matrix product of A^T * A by two methods.
1) "Row times column" as we learned in school.
2) With rearranged loops such that need for cash refreshes is reduced
(this can be improved even more).
Compile: gcc -o aT_a aT_a.c -lgslcblas -lblas -lm
*/
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <cblas.h>
#define ROWS 2000
#define COLS 1000
static double a[ROWS][COLS];
static double c[COLS][COLS];
static void dot() {
int i, j;
double *ai, *bj;
ai = a[0];
for (i=0; i<COLS; i++) {
bj = a[0];
for (j=0; j<COLS; j++) {
c[i][j] = cblas_ddot(ROWS,ai,COLS,bj,COLS);
bj += 1;
}
ai += 1;
}
}
static void axpy() {
int i, j;
double *ci, *bj, aij;
for (i=0; i<COLS; i++) {
ci = c[i];
for (j=0; j<COLS; j++) ci[j] = 0.;
for (j=0; j<ROWS; j++) {
aij = a[j][i];
bj = a[j];
cblas_daxpy(COLS,aij,bj,1,ci,1);
}
}
}
int main(int argc, char** argv) {
clock_t t0, t1;
int i, j;
for (i=0; i<ROWS; ++i)
for (j=0; j<COLS; ++j)
a[i][j] = i+j;
t0 = clock();
dot();
t0 = clock();
printf("Time for DOT : %f sec.\n",(double)t0/CLOCKS_PER_SEC);
axpy();
t1 = clock();
printf("Time for AXPY: %f sec.\n",(double)(t1-t0)/CLOCKS_PER_SEC);
return 0;
}
The CBLAS dot product is effectively just a computation in slightly unrolled loop. The netlib Fortran is just this:
DO I = MP1,N,5
DTEMP = DTEMP + DX(I)*DY(I) + DX(I+1)*DY(I+1) +
$ DX(I+2)*DY(I+2) + DX(I+3)*DY(I+3) + DX(I+4)*DY(I+4)
END DO
ie. just a loop unrolled to a stride of 5.
If you must use a ddot style dot product for your operation, you might get a performance boost by re-writing your loop to use SSE2 intrinsics:
#include <emmintrin.h>
double ddotsse2(const double *x, const double *y, const int n)
{
double result[2];
int n2 = 2 * (n/2);
__m128d dtemp;
if ( (n % 2) == 0) {
dtemp = _mm_setzero_pd();
} else {
dtemp = _mm_set_sd(x[n] * y[n]);
}
for(int i=0; i<n2; i+=2) {
__m128d x1 = _mm_loadr_pd(x+i);
__m128d y1 = _mm_loadr_pd(y+i);
__m128d xy = _mm_mul_pd(x1, y1);
dtemp = _mm_add_pd(dtemp, xy);
}
_mm_store_pd(&result[0],dtemp);
return result[0] + result[1];
}
(not tested, never been compiled, buyer beware).
This may or may be faster than the standard BLAS implementation. You may also want to investigate whether further loop unrolling could improve performance.
If you're not using SSE2 intrinsics or using a data type that may not boost performance with them, you can try to transpose the matrix for an easy improvement in performance for larger matrix multiplications with cblas_?dot. Performing the matrix multiplication in blocks also helps.
void matMulDotProduct(int n, float *A, float* B, int a_size, int b_size, int a_row, int a_col, int b_row, int b_col, float *C) {
int i, j, k;
MKL_INT incx, incy;
incx = 1;
incy = b_size;
//copy out multiplying matrix from larger matrix
float *temp = (float*) malloc(n * n * sizeof(float));
for (i = 0; i < n; ++i) {
cblas_scopy(n, &B[(b_row * b_size) + b_col + i], incy, &temp[i * n], 1);
}
//transpose
mkl_simatcopy('R', 'T', n, n, 1.0, temp, 1, 1);
for (i = 0; i < n; i+= BLOCK_SIZE) {
for (j = 0; j < n; j++) {
for (k = 0; k < BLOCK_SIZE; ++k) {
C[((i + k) * n) + j] = cblas_sdot(n, &A[(a_row + i + k) * a_size + a_col], incx, &temp[n * j], 1);
}
}
}
free(temp);
}
On my machine, this code is about 1 order of magnitude faster than the the 3 loop code (but also 1 order of magnitude slower than cblas_?gemm call) for single precision floats and 2K by 2K matrices. (I'm using Intel MKL).