Understanding Metal usage with AVCaptureSession video output

Question

I'm trying to understand what is the right way to manipulate Video output(CMPixelBuffer), using Metal.

As far as i understand there is MTKView. Each CMPixelBuffer, from the video output, is being assigned to some what of Metal Texture. So the final preview is the from the MTKView?

When i see the final result on the screen, is it :

1)CMSampleBuffer->Metal->CMSampleBuffer

or

2)CMSampleBuffer->Metal->MTKView

Pretty confused. Can someone put things on spot?

Answer 1

The simplest, barebones minimum code required to convert from a sample buffer produced by the video camera to a Metal texture:

@import UIKit;
@import AVFoundation;
@import CoreMedia;
#import <MetalKit/MetalKit.h>
#import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h>

@interface ViewController : UIViewController <MTKViewDelegate, AVCaptureVideoDataOutputSampleBufferDelegate>  {
    NSString *_displayName;
    NSString *serviceType;
}

@property (retain, nonatomic) SessionContainer *session;
@property (retain, nonatomic) AVCaptureSession *avSession;

@end;

#import "ViewController.h"

@interface ViewController () {
    MTKView *_metalView;

    id<MTLDevice> _device;
    id<MTLCommandQueue> _commandQueue;
    id<MTLTexture> _texture;

    CVMetalTextureCacheRef _textureCache;
}

@property (strong, nonatomic) AVCaptureDevice *videoDevice;
@property (nonatomic) dispatch_queue_t sessionQueue;

@end

@implementation ViewController

- (void)viewDidLoad {
    NSLog(@"%s", __PRETTY_FUNCTION__);
    [super viewDidLoad];

    _device = MTLCreateSystemDefaultDevice();
    _metalView = [[MTKView alloc] initWithFrame:self.view.bounds];
    [_metalView setContentMode:UIViewContentModeScaleAspectFit];
    _metalView.device = _device;
    _metalView.delegate = self;
    _metalView.clearColor = MTLClearColorMake(1, 1, 1, 1);
    _metalView.colorPixelFormat = MTLPixelFormatBGRA8Unorm;
    _metalView.framebufferOnly = NO;
    _metalView.autoResizeDrawable = NO;

    CVMetalTextureCacheCreate(NULL, NULL, _device, NULL, &_textureCache);

    [self.view addSubview:_metalView];

    self.sessionQueue = dispatch_queue_create( "session queue", DISPATCH_QUEUE_SERIAL );

    if ([self setupCamera]) {
        [_avSession startRunning];
    }
}

- (BOOL)setupCamera {
    NSLog(@"%s", __PRETTY_FUNCTION__);
    @try {
        NSError * error;

            _avSession = [[AVCaptureSession alloc] init];
            [_avSession beginConfiguration];
            [_avSession setSessionPreset:AVCaptureSessionPreset640x480];

            // get list of devices; connect to front-facing camera
            self.videoDevice = [AVCaptureDevice defaultDeviceWithMediaType:AVMediaTypeVideo];
            if (self.videoDevice == nil) return FALSE;

            AVCaptureDeviceInput *input = [AVCaptureDeviceInput deviceInputWithDevice:self.videoDevice error:&error];
            [_avSession addInput:input];

            dispatch_queue_t sampleBufferQueue = dispatch_queue_create("CameraMulticaster", DISPATCH_QUEUE_SERIAL);

            AVCaptureVideoDataOutput * dataOutput = [[AVCaptureVideoDataOutput alloc] init];
            [dataOutput setAlwaysDiscardsLateVideoFrames:YES];
            [dataOutput setVideoSettings:@{(id)kCVPixelBufferPixelFormatTypeKey: @(kCVPixelFormatType_32BGRA)}];
            [dataOutput setSampleBufferDelegate:self queue:sampleBufferQueue];

            [_avSession addOutput:dataOutput];
            [_avSession commitConfiguration]; 
    } @catch (NSException *exception) {
        NSLog(@"%s - %@", __PRETTY_FUNCTION__, exception.description);
        return FALSE;
    } @finally {
        return TRUE;
    }

}

- (void)captureOutput:(AVCaptureOutput *)captureOutput didOutputSampleBuffer:(CMSampleBufferRef)sampleBuffer fromConnection:(AVCaptureConnection *)connection
{
    CVPixelBufferRef pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
    {
        size_t width = CVPixelBufferGetWidth(pixelBuffer);
        size_t height = CVPixelBufferGetHeight(pixelBuffer);

        CVMetalTextureRef texture = NULL;
        CVReturn status = CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault, _textureCache, pixelBuffer, NULL, MTLPixelFormatBGRA8Unorm, width, height, 0, &texture);
        if(status == kCVReturnSuccess)
        {
            _metalView.drawableSize = CGSizeMake(width, height);
            _texture = CVMetalTextureGetTexture(texture);
            _commandQueue = [_device newCommandQueue];
            CFRelease(texture);
        }
    }
}

- (void)drawInMTKView:(MTKView *)view {
    // creating command encoder
    if (_texture) {
        id<MTLCommandBuffer> commandBuffer = [_commandQueue commandBuffer];
        id<MTLTexture> drawingTexture = view.currentDrawable.texture;

        // set up and encode the filter
        MPSImageGaussianBlur *filter = [[MPSImageGaussianBlur alloc] initWithDevice:_device sigma:5];

        [filter encodeToCommandBuffer:commandBuffer sourceTexture:_texture destinationTexture:drawingTexture];

        // committing the drawing
        [commandBuffer presentDrawable:view.currentDrawable];
        [commandBuffer commit];
        _texture = nil;
    }
}

- (void)mtkView:(MTKView *)view drawableSizeWillChange:(CGSize)size {

}

@end

Answer 2

I'd encourage you to do this in two steps. First, get comfortable using Metal and MTKView to draw something, anything, to the screen. Second, after you know how to draw with Metal (create and encode to command buffers, present drawables, etc), you can apply the texture you generate from the CMSampleBuffer to a full-screen quad or whatever other geometry you want.

There's a camera capture sample by @McZonk available here that implements a simple example of this. It uses a two-plane YCbCr sample buffer, but if available, you can also request a BGRA formatted sample buffer that can then be converted into a single MTLTexture with CVMetalTextureCacheCreateTextureFromImage and sampled in a shader without further manual conversion.