I'm doing some image processing with opencv::cuda so what I end up with is a cv::cuda::GpuMat. I now want to encode it using ffmpeg(so I can choose the encoder to be hardware accelerated or not). Now I wonder if i can somehow keep the data on the GPU for the encoder without downloading it, because that seems to be the bottleneck in my application running multiple threads.
I'm resizing the images with Opencv CUDA so I have less to download. (resizing with sws_scale makes no difference)
cv::cuda::GpuMat currentFrame;
...
cv::cuda::GpuMat resized;
cv::cuda::resize(currentFrame,resized,cv::Size(width*0.75,height*0.75),0,0,cv::INTER_NEAREST);
cv::Mat frameEnc = cv::Mat(resized);
const int stride[] = { static_cast<int>(frameEnc.step[0]) };
sws_scale(swsctx, &frameEnc.data, stride, 0, frameEnc.rows, avframe->data, avframe->linesize);
ret = avcodec_send_frame(codec, avframe);
if(!ret) {
/* rescale packet timestamp */
pkt->duration = 1;
av_packet_rescale_ts(pkt, codec->time_base, vstrm->time_base);
/* write packet */
av_write_frame(outctx, pkt);
}
Now this does work and performs ok, but I really wish I could do something like:
cv::cuda::GpuMat currentFrame;
...
GpuMatToAvFrame(currentFrame,avframe);
ret = avcodec_send_frame(codec, avframe);
if(!ret) {
/* rescale packet timestamp */
pkt->duration = 1;
av_packet_rescale_ts(pkt, codec->time_base, vstrm->time_base);
/* write packet */
av_write_frame(outctx, pkt);
}
where the avframe data is also on the gpu so that I don't download need any transfer between GPU-CPU/CPU-GPU
I think the class cv::cudacodec::VideoWriter could help, once an issue with OpenCV gets fixed. The class allows you to write a GpuMat directly. However I believe that due to a bug in OpenCV, you can't build OpenCV with support for this class. Which means this isn't a great solution now, but might be in the future.
Related
I'm using an STM32 (STM32F446RE) to receive audio from two INMP441 mems microphone in an stereo setup via I2S protocol and record it into a .WAV on a micro SD card, using the HAL library.
I wrote the firmware that records audio into a .WAV with FreeRTOS. But the audio files that I record sound like Darth Vader. Here is a screenshot of the audio in audacity:
if you zoom in you can see a constant noise being inserted in between the real audio data:
I don't know what is causing this.
I have tried increasing the MessageQueue, but that doesnt seem to be the problem, the queue is kept at 0 most of the time. I've tried different frame sizes and sampling rates, changing the number of channels, using only one inmp441. All this without any success.
I proceed explaining the firmware.
Here is a block diagram of the architecture for the RTOS that I have implemented:
It consists of three tasks. The first one receives a command via UART (with interrupts) that signals to start or stop recording. the second one is simply an state machine that walks through the steps to write a .WAV.
Here the code for the WriteWavFileTask:
switch(audio_state)
{
case STATE_START_RECORDING:
sprintf(filename, "%saud_%03d.wav", SDPath, count++);
do
{
res = f_open(&file_ptr, filename, FA_CREATE_ALWAYS|FA_WRITE);
}
while(res != FR_OK);
res = fwrite_wav_header(&file_ptr, I2S_SAMPLE_FREQUENCY, I2S_FRAME, 2);
HAL_I2S_Receive_DMA(&hi2s2, aud_buf, READ_SIZE);
audio_state = STATE_RECORDING;
break;
case STATE_RECORDING:
osDelay(50);
break;
case STATE_STOP:
HAL_I2S_DMAStop(&hi2s2);
while(osMessageQueueGetCount(AudioQueueHandle)) osDelay(1000);
filesize = f_size(&file_ptr);
data_len = filesize - 44;
total_len = filesize - 8;
f_lseek(&file_ptr, 4);
f_write(&file_ptr, (uint8_t*)&total_len, 4, bw);
f_lseek(&file_ptr, 40);
f_write(&file_ptr, (uint8_t*)&data_len, 4, bw);
f_close(&file_ptr);
audio_state = STATE_IDLE;
break;
case STATE_IDLE:
osThreadSuspend(WAVHandle);
audio_state = STATE_START_RECORDING;
break;
default:
osDelay(50);
break;
Here are the macros used in the code for readability:
#define I2S_DATA_WORD_LENGTH (24) // industry-standard 24-bit I2S
#define I2S_FRAME (32) // bits per sample
#define READ_SIZE (128) // samples to read from I2S
#define WRITE_SIZE (READ_SIZE*I2S_FRAME/16) // half words to write
#define WRITE_SIZE_BYTES (WRITE_SIZE*2) // bytes to write
#define I2S_SAMPLE_FREQUENCY (16000) // sample frequency
The last task is the responsible for processing the buffer received via I2S. Here is the code:
void convert_endianness(uint32_t *array, uint16_t Size) {
for (int i = 0; i < Size; i++) {
array[i] = __REV(array[i]);
}
}
void HAL_I2S_RxCpltCallback(I2S_HandleTypeDef *hi2s)
{
convert_endianness((uint32_t *)aud_buf, READ_SIZE);
osMessageQueuePut(AudioQueueHandle, aud_buf, 0L, 0);
HAL_I2S_Receive_DMA(hi2s, aud_buf, READ_SIZE);
}
void pvrWriteAudioTask(void *argument)
{
/* USER CODE BEGIN pvrWriteAudioTask */
static UINT *bw;
static uint16_t aud_ptr[WRITE_SIZE];
/* Infinite loop */
for(;;)
{
osMessageQueueGet(AudioQueueHandle, aud_ptr, 0L, osWaitForever);
res = f_write(&file_ptr, aud_ptr, WRITE_SIZE_BYTES, bw);
}
/* USER CODE END pvrWriteAudioTask */
}
This tasks reads from a queue an array of 256 uint16_t elements containing the raw audio data in PCM. f_write takes the Size parameter in number of bytes to write to the SD card, so 512 bytes. The I2S Receives 128 frames (for a 32 bit frame, 128 words).
The following is the configuration for the I2S and clocks:
Any help would be much appreciated!
Solution
As pmacfarlane pointed out, the problem was with the method used for buffering the audio data. The solution consisted of easing the overhead on the ISR and implementing a circular DMA for double buffering. Here is the code:
#define I2S_DATA_WORD_LENGTH (24) // industry-standard 24-bit I2S
#define I2S_FRAME (32) // bits per sample
#define READ_SIZE (128) // samples to read from I2S
#define BUFFER_SIZE (READ_SIZE*I2S_FRAME/16) // number of uint16_t elements expected
#define WRITE_SIZE_BYTES (BUFFER_SIZE*2) // bytes to write
#define I2S_SAMPLE_FREQUENCY (16000) // sample frequency
uint16_t aud_buf[2*BUFFER_SIZE]; // Double buffering
static volatile int16_t *BufPtr;
void convert_endianness(uint32_t *array, uint16_t Size) {
for (int i = 0; i < Size; i++) {
array[i] = __REV(array[i]);
}
}
void HAL_I2S_RxHalfCpltCallback(I2S_HandleTypeDef *hi2s)
{
BufPtr = aud_buf;
osSemaphoreRelease(RxAudioSemHandle);
}
void HAL_I2S_RxCpltCallback(I2S_HandleTypeDef *hi2s)
{
BufPtr = &aud_buf[BUFFER_SIZE];
osSemaphoreRelease(RxAudioSemHandle);
}
void pvrWriteAudioTask(void *argument)
{
/* USER CODE BEGIN pvrWriteAudioTask */
static UINT *bw;
/* Infinite loop */
for(;;)
{
osSemaphoreAcquire(RxAudioSemHandle, osWaitForever);
convert_endianness((uint32_t *)BufPtr, READ_SIZE);
res = f_write(&file_ptr, BufPtr, WRITE_SIZE_BYTES, bw);
}
/* USER CODE END pvrWriteAudioTask */
}
Problems
I think the problem is your method of buffering the audio data - mainly in this function:
void HAL_I2S_RxCpltCallback(I2S_HandleTypeDef *hi2s)
{
convert_endianness((uint32_t *)aud_buf, READ_SIZE);
osMessageQueuePut(AudioQueueHandle, aud_buf, 0L, 0);
HAL_I2S_Receive_DMA(hi2s, aud_buf, READ_SIZE);
}
The main problem is that you are re-using the same buffer each time. You have queued a message to save aud_buf to the SD-card, but you've also instructed the I2S to start DMAing data into that same buffer, before it has been saved. You'll end up saving some kind of mish-mash of "old" data and "new" data.
#Flexz pointed out that the message queue takes a copy of the data, so there is no issue about the I2S writing over the data that is being written to the SD-card. However, taking the copy (in an ISR) adds overhead, and delays the start of the new I2S DMA.
Another problem is that you are doing the endian conversion in this function (that is called from an ISR). This will block any other (lower priority) interrupts from being serviced while this happens, which is a bad thing in an embedded system. You should do the endian conversion in the task that reads from the queue. ISRs should be very short and do the minimum possible work (often just setting a flag, giving a semaphore, or adding something to a queue).
Lastly, while you are doing the endian conversion, what is happening to audio samples? The previous DMA has completed, and you haven't started a new one, so they will just be dropped on the floor.
Possible solution
You probably want to allocate a suitably big buffer, and configure your DMA to work in circular buffer mode. This means that once started, the DMA will continue forever (until you stop it), so you'll never drop any samples. There won't be any gap between one DMA finishing and a new one starting, since you never need to start a new one.
The DMA provides a "half-complete" interrupt, to say when it has filled half the buffer. So start the DMA, and when you get the half-complete interrupt, queue up the first half of the buffer to be saved. When you get the fully-complete interrupt, queue up the second half of the buffer to be saved. Rinse and repeat.
You might want to add some logic to detect if the interrupt happens before the previous save has completed, since the data will be overrun and possibly corrupted. Depending on the speed of the SD-card (and the sample rate), this may or may not be a problem.
I am trying to write a code that uses opencv Mat objects it goes something like this
Mat img;
vector<Mat> images;
for (i = 1; i < 5; i++)
{
img.create(h,w,type) // h,w and type are given correctly
// input an image from somewhere to img correctly.
images.push_back(img);
img.release()
}
for (i = 1; i < 5; i++)
images[i].release();
I however still seem to have memory leakage can anyone tell me why it is so?
I thought that if the refcount of a mat object = 0 then the memory should be automatically deallocated
You rarely need to call release explicitly, since OpenCV Mat objects take automatically care of internal memory.
Also take care that Mat copy just copies creates a new header pointing to the same data. If the original Mat goes out of scope you are left with an invalid matrix. So when you push the image into the vector, use a deep copy (clone()) to avoid that it the image into the vector becomes invalid.
Since you mentioned:
I have a large 3D image stored in a Mat object. I am running over it using for loops. creating a 2D mat called "image" putting the slices into image, pushing back image to vector images. releasing the image. And later doing a for loop on the images vector releasing all the matrices one by one.
You can store all slices into the vector with the following code. To release the images in the vector, just clear the vector.
#include <opencv2/opencv.hpp>
#include <vector>
using namespace cv;
using namespace std;
int main()
{
// Init the multidimensional image
int sizes[] = { 10, 7, 5 };
Mat data(3, sizes, CV_32F);
randu(data, Scalar(0, 0, 0), Scalar(1,1,1));
// Put slices into images
vector<Mat> images;
for (int z = 0; z < data.size[2]; ++z)
{
// Create the slice
Range ranges[] = { Range::all(), Range::all(), Range(z, z + 1) };
Mat slice(data(ranges).clone()); // with clone slice is continuous, but still 3d
Mat slice2d(2, &data.size[0], data.type(), slice.data); // make the slice a 2d image
// Clone the slice into the vector, or it becomes invalid when slice goes of of scope.
images.push_back(slice2d.clone());
}
// You can deallocate the multidimensional matrix now, if needed
data.release();
// Work with slices....
// Release the vector of slices
images.clear();
return 0;
}
Please try this code, which is basically what you do:
void testFunction()
{
// image width/height => 80MB images
int size = 5000;
cv::Mat img = cv::Mat(size, size, CV_8UC3);
std::vector<cv::Mat> images;
for (int i = 0; i < 5; i++)
{
// since image size is the same for i==0 as the initial image, no new data will be allocated in the first iteration.
img.create(size+i,size+i,img.type()); // h,w and type are given correctly
// input an image from somewhere to img correctly.
images.push_back(img);
// release the created image.
img.release();
}
// instead of manual releasing, a images.clear() would have been enough here.
for(int i = 0; i < images.size(); i++)
images[i].release();
images.clear();
}
int main()
{
cv::namedWindow("bla");
cv::waitKey(0);
for(unsigned int i=0; i<100; ++i)
{
testFunction();
std::cout << "another iteration finished" << std::endl;
cv::waitKey(0);
}
std::cout << "end of main" << std::endl;
cv::waitKey(0);
return 0;
}
After the first call of testFunction, memory will be "leaked" so that the application consumes 4 KB more memory on my device. But not more "leaks" after additional calls for me...
So this looks like your code is ok and the "memory leak" isn't related to that matrix creation and releasing, but maybe to some "global" things happening within the openCV library or C++ to optimize future function calls or memory allocations.
I encountered same problems when iterate openCV mat. The memory consumption can be 1.1G, then it stopped by warning that no memory. In my program, there are macro #define new new(FILE, LINE), crashed with some std lib. So I deleted all Overloading Operator about new/delete. When debugging, it has no error. But when it runs, I got "Debug Assertion Failed! Expression: _pFirstBlock == pHead". Following the instruction
Debug Assertion Error in OpenCV
I changed setting from MT (Release)/MTd (Debug)to
Project Properties >> Configuration Properties >> C/C++ >> Code Generation and changing the Runtime Library to:
Multi-threaded Debug DLL (/MDd), if you are building the Debug version of your code.
Multi-threaded DLL(/MD), if you are building the Release version of your code.
The memory leak is gone. The memory consumption is 38M.
I would like to use the EZAudio framework to do realtime microphone signal FFT processing, along with some other processing in order to determine the peak frequency.
The problem is, the EZmicrophone class only appears to work on 512 samples, however, my signal requires an FFT of 8192 or even 16384 samples. There doesnt appear to be a way to change the buffer size in EZMicrophone, but I've read posts that recommend creating an array of my target size and appending the microphone buffer to it, then when it's full, do the FFT.
When I do this though, I get large chunks of memory with no data, or discontinuities between the segments of copied memory. I think it may have something to do with the timing or order in which the microphone delegate is being called or memory being overwritten in different threads...I'm grasping at straws here. Am I correct in assuming that this code is being executed everytime the microphone buffer is full of a new 512 samples?
Can anyone suggest what I may be doing wrong? I've been stuck on this for a long time.
Here is the post I've been using as a reference:
EZAudio: How do you separate the buffersize from the FFT window size(desire higher frequency bin resolution).
// Global variables which are bad but I'm just trying to make things work
float tempBuf[512];
float fftBuf[8192];
int samplesRemaining = 8192;
int samplestoCopy = 512;
int FFTLEN = 8192;
int fftBufIndex = 0;
#pragma mark - EZMicrophoneDelegate
-(void) microphone:(EZMicrophone *)microphone
hasAudioReceived:(float **)buffer
withBufferSize:(UInt32)bufferSize
withNumberOfChannels:(UInt32)numberOfChannels {
// Copy the microphone buffer so it wont be changed
memcpy(tempBuf, buffer[0], bufferSize);
dispatch_async(dispatch_get_main_queue(),^{
// Setup the FFT if it's not already setup
if( !_isFFTSetup ){
[self createFFTWithBufferSize:FFTLEN withAudioData:fftBuf];
_isFFTSetup = YES;
}
int samplesRemaining = FFTLEN;
memcpy(fftBuf+fftBufIndex, tempBuf, samplestoCopy*sizeof(float));
fftBufIndex += samplestoCopy;
samplesRemaining -= samplestoCopy;
if (fftBufIndex == FFTLEN)
{
fftBufIndex = 0;
samplesRemaining = FFTLEN;
[self updateFFTWithBufferSize:FFTLEN withAudioData:fftBuf];
}
});
}
You likely have threading issues because you are trying to do work in some blocks that takes much much longer than the time between audio callbacks. Your code is being called repeatedly before prior calls can say that they are done (with the FFT setup or clearing the FFT buffer).
Try doing the FFT setup outside the callback before starting the recording, only copy to a circular buffer or FIFO inside the callback, and do the FFT in code async to the callback (not locked in the same block as the circular buffer copy).
I'm working on an audio visualizer in C with OpenGL, Libsamplerate, portaudio, and libsndfile. I'm having difficulty using src_process correctly within my whole paradigm. My goal is to use src_process to achieve Vinyl Like varispeed in real time within the visualizer. Right now my implementation changes the pitch of the audio without changing the speed. It does so with lots of distortion due to what sounds like missing frames as when I lower the speed with the src_ratio it almost sounds granular like chopped up samples. Any help would be appreciated, I keep experimenting with my buffering chunks however 9 times out of 10 I get a libsamplerate error saying my input and output arrays are overlapping. I've also been looking at the speed change example that came with libsamplerate and I can't find where I went wrong. Any help would be appreciated.
Here's the code I believe is relevant. Thanks and let me know if I can be more specific, this semester was my first experience in C and programming.
#define FRAMES_PER_BUFFER 1024
#define ITEMS_PER_BUFFER (FRAMES_PER_BUFFER * 2)
float src_inBuffer[ITEMS_PER_BUFFER];
float src_outBuffer[ITEMS_PER_BUFFER];
void initialize_SRC_DATA()
{
data.src_ratio = 1; //Sets Default Playback Speed
/*---------------*/
data.src_data.data_in = data.src_inBuffer; //Point to SRC inBuffer
data.src_data.data_out = data.src_outBuffer; //Point to SRC OutBuffer
data.src_data.input_frames = 0; //Start with Zero to Force Load
data.src_data.output_frames = ITEMS_PER_BUFFER
/ data.sfinfo1.channels; //Number of Frames to Write Out
data.src_data.src_ratio = data.src_ratio; //Sets Default Playback Speed
}
/* Open audio stream */
err = Pa_OpenStream( &g_stream,
NULL,
&outputParameters,
data.sfinfo1.samplerate,
FRAMES_PER_BUFFER,
paNoFlag,
paCallback,
&data );
/* Read FramesPerBuffer Amount of Data from inFile into buffer[] */
numberOfFrames = sf_readf_float(data->inFile, data->src_inBuffer, framesPerBuffer);
/* Looping of inFile if EOF is Reached */
if (numberOfFrames < framesPerBuffer)
{
sf_seek(data->inFile, 0, SEEK_SET);
numberOfFrames = sf_readf_float(data->inFile,
data->src_inBuffer+(numberOfFrames*data->sfinfo1.channels),
framesPerBuffer-numberOfFrames);
}
/* Inform SRC Data How Many Input Frames To Process */
data->src_data.end_of_input = 0;
data->src_data.input_frames = numberOfFrames;
/* Perform SRC Modulation, Processed Samples are in src_outBuffer[] */
if ((data->src_error = src_process (data->src_state, &data->src_data))) {
printf ("\nError : %s\n\n", src_strerror (data->src_error)) ;
exit (1);
}
* Write Processed SRC Data to Audio Out and Visual Out */
for (i = 0; i < framesPerBuffer * data->sfinfo1.channels; i++)
{
// gl_audioBuffer[i] = data->src_outBuffer[i] * data->amplitude;
out[i] = data->src_outBuffer[i] * data->amplitude;
}
I figured out a solution that works well enough for me and am just going to explain it best I can for anyone else with a similar issue. So to get the Varispeed to work, the way the API works is you give it a certain number of frames, and it spits out a certain number of frames. So for a SRC ratio of 0.5, if you process 512 frames per loop you are feeding in 512/0.5 frames = 1024 frames. That way when the API runs its src_process function, it compresses those 1024 frames into 512, speeding up the samples. So I dont fully understand why it solved my issue, but the problem was if the ratio is say 0.7, you end up with a float number which doesn't work with the arrays indexed int values. Therefore there's missing samples unless the src ratio is eqaully divisble by the framesperbuffer potentially at the end of each block. So what I did was add +2 frames to be read if the framesperbuffer%src.ratio != 0 and it seemed to fix 99% of the glitches.
/* This if Statement Ensures Smooth VariSpeed Output */
if (fmod((double)framesPerBuffer, data->src_data.src_ratio) == 0)
{
numInFrames = framesPerBuffer;
}
else
numInFrames = (framesPerBuffer/data->src_data.src_ratio) + 2;
/* Read FramesPerBuffer Amount of Data from inFile into buffer[] */
numberOfFrames = sf_readf_float(data->inFile, data->src_inBuffer, numInFrames);
I am using ffmpeg library. I want to know how much memory one packet can take.
I debug to check the members in on AVPacket, and none of them seem reasonable, such as AVPacket.size, ec.
If you provide your own data buffer, it needs to have a size of mininum FF_MIN_BUFFER_SIZE. You would then set the AVPacket.size to the allocated size, and AVPacket.data to the memory you've allocated.
Note that all FFmpeg decoding routine will simply fail if you provide your own buffer and it's too small.
The other possibility, is let FFmpeg calculates the optimal size for you.
Then do something like:
AVPacket pkt;
pkt.size = 0;
pkt.data = NULL; // <-- the critical part is there
int got_output = 0;
ret = avcodec_encode_audio2(ctx, &pkt, NULL, &got_output);
and provide this AVPacket to the encoding codec. Memory will be allocated automatically.
You will have to call av_free_packet upon return from the encoder and if got_output is set to 1.
FFmpeg will automatically free the AVPacket content in case of error.
AVPacket::size holds the size of the referenced data. Because it is a generic container for data, there can be no definite answer to the question
how much memory one packet can take
It can actually take from zero to a lot. Everything depends on data type, codec and other related parameters.
From FFmpeg examples:
static void audio_encode_example(const char *filename)
{
// ...
AVPacket pkt;
// ...
ret = avcodec_encode_audio2(c, &pkt, NULL, &got_output);
// ...
if (got_output) {
fwrite(pkt.data, 1, pkt.size, f); // <<--- AVPacket.size
av_free_packet(&pkt);
}