Explore Live GPU Profiling with Metal Counters

3:38 - Checking for available Metal counters

if (@available(macOS 11.0, iOS 14.0, *))
{
    _supportsStageBoundary = [_device supportsCounterSampling:MTLCounterSamplingPointAtStageBoundary];
    _supportsDrawBoundary  = [_device supportsCounterSampling:MTLCounterSamplingPointAtDrawBoundary];
}

3:52 - Counter sets

[_device.counterSets enumerateObjectsUsingBlock:^(id<MTLCounterSet> nonnull obj,
                                                  NSUInteger                idx,
                                                  BOOL * nonnull            stop) {
       if ([[obj name] isEqualToString:MTLCommonCounterSetTimestamp])
            _counterSetTimestamp = obj;
}];

5:05 - Sampling counters on Apple GPUs


// When setting up the render pass descriptor

if (_supportsStageBoundary || _supportsDrawBoundary)
{
    MTLCounterSampleBufferDescriptor *desc = [MTLCounterSampleBufferDescriptor new];

    desc.sampleCount = 6; // Number of samples to store 
    desc.storageMode = MTLStorageModeShared;
    desc.label       = @"Live Profiling HUD Metal counter sample buffer";
    desc.counterSet  = _counterSetTimestamp;

    id<MTLCounterSampleBuffer> sampleBuffer =
                               [_device newCounterSampleBufferWithDescriptor:desc error:nil];

    MTLRenderPassSampleBufferAttachmentDescriptor *sampleBufferDesc =
                                  renderPassDescriptor.sampleBufferAttachments[0];

    if (_supportsStageBoundary)
    {
        sampleBufferDesc.startOfVertexSampleIndex   = 0;
        sampleBufferDesc.endOfVertexSampleIndex     = 1;
        sampleBufferDesc.startOfFragmentSampleIndex = 2;
        sampleBufferDesc.endOfFragmentSampleIndex   = 3;
    }

    sampleBufferDesc.sampleBuffer = sampleBuffer;
}

6:23 - Sampling counters at draw boundary

// After creating a new render command encoder
[renderCommandEncoder sampleCountersInBuffer:sampleBuffer
                               atSampleIndex:4
                                 withBarrier:NO];

// All draw calls
[renderCommandEncoder sampleCountersInBuffer:sampleBuffer
                               atSampleIndex:5
                                 withBarrier:NO];

// End encoding

7:28 - Collecting timestamps

// For each tracked sampleBuffer, resolve the counters
NSData *data = [sampleBuffer resolveCounterRange:NSMakeRange(0, 6)];

MTLCounterResultTimestamp *sample = (MTLCounterResultTimestamp *)[data bytes];

// And simply access the timestamps
if (_supportsStageBoundary)
{
    double vertexStart = sample[0].timestamp / (double)NSEC_PER_SEC;
}

// Check for errors
if (sample[0].timestamp == MTLCounterErrorValue) 
{
  // Handle error
}

9:05 - Aligning timestamps

// On immediate mode GPU
MTLTimestamp cpuTimestamp;
MTLTimestamp gpuTimestamp;
[_device sampleTimestamps:&cpuTimestamp gpuTimestamp:&gpuTimestamp];

// Do a linear interpolation between correlated timestamps
gpu_ns = cpu_t0 + (cpu_t1 - cpu_t0) * (gpu_timestamp - gpu_t0) / (gpu_t1 - gpu_t0);