Develop Reference

What is the most performant method for setting a subset of sequential items in a large array?

I am working with camera streams. I bring in 1,228,800 bytes per frame, so efficiency is crucial and nanoseconds per byte add up quickly.
I've come up with some example code below to describe the problem as succinctly as possible without seeming too contrived.
There are plenty of inefficiencies in this example code such as defining variables inside of loops, or dividing the brightness value instead of just using a composite value. These aren't my concern, and are just there to make the example simpler.
What I need advice on is the most performant method in C# for setting 3 sequential values at some determined location in a very large array, such as in the case below where I'm setting BGR to 255 while skipping the 4th byte.
Edit: To clarify, the concerning issue is where I'm reindexing Output for each index that is being set. It seems like there should potentially be some method for not traversing the entire array for each value if I already have the location of the previous item.
// Colors are stored as 4 bytes: BGRX where X is always 0
public byte[] Input = new byte[640 * 480 * 4];
public byte[] Output = new byte[640 * 480 * 4];
public int Threshold = 180;
public void ProcessImage() {
for (int i = 0; i < Input.Length; i += 4) {
var brightness = (Input[i] + Input[i + 1] + Input[i + 2]) / 3; // some value under 255
if (brightness > Threshold) {
// What is the most efficient way possible to do this?
Output[i] = 255 - Input[i];
Output[i + 1] = 255 - Input[i + 1];
Output[i + 2] = 255 - Input[i + 2];
}
else {
Output[i] = Input[i];
Output[i + 1] = Input[i + 1];
Output[i + 2] = Input[i + 2];
}
}
}

This (untested, and unsafe) code should be faster, if all you care about is speed:
public void ProcessImage()
{
int ilength = Input.Length;
Debug.Assert(ilength == Output.Length);
Debug.Assert(ilength%4 == 0);
unsafe
{
GCHandle pinned1 = GCHandle.Alloc(Input, GCHandleType.Pinned);
byte* input = (byte*)pinned1.AddrOfPinnedObject();
GCHandle pinned2 = GCHandle.Alloc(Input, GCHandleType.Pinned);
byte* output = (byte*)pinned2.AddrOfPinnedObject();
for (int i = 0; i < ilength; i += 4)
{
var brightness = (*(input) + *(input + 1) + *(input + 2)) / 3;
if (brightness > Threshold)
{
// What is the most efficient way possible to do this?
(*(output)) = (byte)(255 - *(input));
(*(output+1)) = (byte)(255 - *(input+1));
(*(output+2)) = (byte)(255 - *(input+2));
}
else
{
(*(output)) = *(input);
(*(output + 1)) = *(input + 1);
(*(output + 2)) = *(input + 2);
}
input += 4;
output += 4;
}
pinned1.Free();
pinned2.Free();
}
}
Note that I've incorporate the necessary assumptions at the top of the function. I'd suggest you always do this, but whether you prefer Debug.Assert or some other form of validation is up to you.

If you're happy to carry the 4th byte through, it would be quicker to copy Input to Output first with a block copy, then not to perform the else clause of the branch:
Buffer.BlockCopy(Input,0,Output,0,Input.Length);
for (int i = 0; i < Input.Length; i += 4) {
var brightness = (Input[i] + Input[i + 1] + Input[i + 2]) / 3;
if (brightness > Threshold) {
Output[i] = (byte)(255 - Input[i]);
Output[i + 1] = (byte)(255 - Input[i + 1]);
Output[i + 2] = (byte)(255 - Input[i + 2]);
}
}

In terms of the most performant way of setting a single value to multiple array indicies in c#, I think you're looking at it. There's no non-looping way to set the same value to multiple indicies. See How can I assign a value to multiple array indices at once without looping?
If it helps, there's no need for the else statement where you set the 3 indicies to 0. default(byte) is already zero, so every index in the Ouput[] array will initialize to 0.
As a side note, defining variables inside of loops vs outside of loops has no effect on the resulting IL. See Is it better to declare a variable inside or outside a loop?
EDIT: To add on to the comment above, you can use unsafe methods. See https://stackoverflow.com/a/5375552/3290789 and http://www.gutgames.com/post/Using-Unsafe-Code-for-Faster-Image-Manipulation.aspx

Resample loopback capture

I successfully captured sound from Wasapi using the following code:
IWaveIn waveIn = new WasapiLoopbackCapture();
waveIn.DataAvailable += OnDataReceivedFromWaveOut;
What I need to do now, is to resample the in-memory data to pcm with a sample rate of 8000 and 16 bits per sample mono.
I can't use ACMStream to resample the example, because the recorded audio is 32 bits per second.
I have tried this code to convert bytes from 32 to 16 bits, but all I get every time is just blank audio.
byte[] newArray16Bit = new byte[e.BytesRecorded / 2];
short two;
float value;
for (int i = 0, j = 0; i < e.BytesRecorded; i += 4, j += 2)
{
value = (BitConverter.ToSingle(e.Buffer, i));
two = (short)(value * short.MaxValue);
newArray16Bit[j] = (byte)(two & 0xFF);
newArray16Bit[j + 1] = (byte)((two >> 8) & 0xFF);
}
source = newArray16Bit;

I use this routine to resample on the fly from WASAPI IeeeFloat to the format I need in my app, which is 16kHz, 16 bit, 1 channel. My formats are fixed, so I'm hardcoding the conversions I need, but it can be adapted as needed.
private void ResampleWasapi(object sender, WaveInEventArgs e)
{
//the result of downsampling
var resampled = new byte[e.BytesRecorded / 12];
var indexResampled = 0;
//a variable to flag the mod 3-ness of the current sample
var arity = -1;
var runningSamples = new short[3];
for(var offset = 0; offset < e.BytesRecorded; offset += 8)
{
var float1 = BitConverter.ToSingle(e.Buffer, offset);
var float2 = BitConverter.ToSingle(e.Buffer, offset + 4);
//simple average to collapse 2 channels into 1
float mono = (float)((double)float1 + (double)float2) / 2;
//convert (-1, 1) range int to short
short sixteenbit = (short)(mono * 32767);
//the input is 48000Hz and the output is 16000Hz, so we need 1/3rd of the data points
//so save up 3 running samples and then mix and write to the file
arity = (arity + 1) % 3;
//record the value
runningSamples[arity] = sixteenbit;
//if we've hit the third one
if (arity == 2)
{
//simple average of the 3 and put in the 0th position
runningSamples[0] = (short)(((int)runningSamples[0] + (int)runningSamples[1] + (int)runningSamples[2]) / 3);
//copy that short (2 bytes) into the result array at the current location
Buffer.BlockCopy(runningSamples, 0, resampled, indexResampled, 2);
//for next pass
indexResampled += 2;
}
}
//and tell listeners that we've got data
RaiseDataEvent(resampled, resampled.Length);
}

Unexpected alpha values in DXT5 decompression

I am decompressing an image compressed with DXT5. According to the description each block of 16 bytes starts with 2 bytes alpha data. If i have a look at my file in a hex editor i find that 90% of the image have an alpha value of less than 0.04 (value in the file is < 10) which should not be the case.
If i render the image with OpenGL and let glCompressedTexImage do the work it looks ok. With my code the image is transparent as i would have expected from those small alpha values. The code i use to generate the alpha values looks like that:
byte alpha1 = reader.ReadByte();
byte alpha2 = reader.ReadByte();
uint[] alphaValues = new uint[8]
{
alpha1,
alpha2,
0, 0, 0, 0, 0, 0
};
if (alpha1 > alpha2)
{
for (int i = 0; i < 6; ++i)
{
byte value = (byte)(((6.0f - i) * alpha1 + (1.0f + i) * alpha2) / 7.0f);
alphaValues[i + 2] = value;
}
}
else
{
for (int i = 0; i < 4; ++i)
{
byte value = (byte)(((4.0f - i) * alpha1 + (1.0f + i) * alpha2) / 5.0f);
alphaValues[i + 2] = value;
}
alphaValues[6] = 0;
alphaValues[7] = 255;
}
alpha1 and alpha2 usually are the same (values are like 8 or 3 or 9, the maximum alpha value in the image however is 96).
The colors however are ok. If i render the image without alpha values it looks perfect. Enabling alpha -> transparent.

C# - How to convert 10 bytes to unsigned long

I have 10 bytes - 4 bytes of low order, 4 bytes of high order, 2 bytes of highest order - that I need to convert to an unsigned long. I've tried a couple different methods but neither of them worked:
Try #1:
var id = BitConverter.ToUInt64(buffer, 0);
Try #2:
var id = GetID(buffer, 0);
long GetID(byte[] buffer, int startIndex)
{
var lowOrderUnitId = BitConverter.ToUInt32(buffer, startIndex);
var highOrderUnitId = BitConverter.ToUInt32(buffer, startIndex + 4);
var highestOrderUnitId = BitConverter.ToUInt16(buffer, startIndex + 8);
return lowOrderUnitId + (highOrderUnitId * 100000000) + (highestOrderUnitId * 10000000000000000);
}
Any help would be appreciated, thanks!

As the comments indicate, 10 bytes will not fit in a long (which is a 64-bit data type - 8 bytes). However, you could use a decimal (which is 128-bits wide - 16 bytes):
var lowOrderUnitId = BitConverter.ToUInt32(buffer, startIndex);
var highOrderUnitId = BitConverter.ToUInt32(buffer, startIndex + 4);
var highestOrderUnitId = BitConverter.ToUInt16(buffer, startIndex + 8);
decimal n = highestOrderUnitId;
n *= UInt32.MaxValue;
n += highOrderUnitId;
n *= UInt32.MaxValue;
n += lowOrderUnitId;
I've not actually tested this, but I think it will work...

As has been mentioned, a ulong isn't large enough to hold 10 bytes of data, it's only 8 bytes. You'd need to use a Decimal. The most efficient way (not to mention least code) would probably be to get a UInt64 out of it first, then add the high-order bits:
ushort high = BitConverter.ToUInt16(buffer, 0);
ulong low = BitConverter.ToUInt64(buffer, 2);
decimal num = (decimal)high * ulong.MaxValue + high + low;
(You need to add high a second time because otherwise you'd need to multiply by the value ulong.MaxValue + 1, and that's a lot of annoying casting and parentheses.)

Calculating RMS with overlapping windows

I have a 1-dimensional float array of root mean square values, each calculated with the same window length. Let's say
RMS = {0, 0.01, 0.4, ... }
Now the RMS for a larger window, which can be represented as a range of the original windows, can be calculated as the RMS of the "participating" RMS values from RMS[i] to RMS[i + len]. Here len is the length of the larger window divided by the lenght of the original windows.
I'd like to create a rolling window. I want
rollingRMS[0] = RMS from 0 to len
...
rollingRMS[n] = RMS from n to len+n
calculated as efficiently as possible. I know this isn't very hard to crack, but does anyone have ready code for this?
EDIT: I asked for sample code, so I guess it would be decent to provide some. The following is based on pierr's answer and is written in C#. It's a bit different from my original question as I realized it would be nice to have the resulting array to have the same size as the original and to have the windows end at each element.
// The RMS data to be analysed
float[] RMS = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
// The resulting rolling RMS values
float[] rollingRMS = new float[RMS.Length];
// Window lenght
int len = 3;
// Calculate: rollingRMS will hold root mean square from windows which end at
// each respective sample in the RMS array. For the first len samples the input
// will be treated as zero-padded
for (int i = 0; i < RMS.Length; i++)
{
if (i == 0)
rollingRMS[i] = (float)Math.Sqrt((RMS[i] * RMS[i] / len));
else if (i < len)
rollingRMS[i] = (float)Math.Sqrt(
( RMS[i] * RMS[i] +
len * (rollingRMS[i - 1] * rollingRMS[i - 1])
) / len);
else
rollingRMS[i] = (float)Math.Sqrt(
( len * (rollingRMS[i - 1] * rollingRMS[i - 1]) +
RMS[i] * RMS[i] -
RMS[i - len] * RMS[i - len]
) / len);
}

I am not sure that I have understood your problem correctly. But let me have a try.
a=[1,2,3,4,5,6,7,8,9,10]
LEN = 3
SquareOfRollingRMS[0] = (a[0]^2 + a[1]^2 + a[2]^2 ) / LEN
SquareOfRollingRMS[1] = ( a[1]^2 + a[2]^2 + a[3]^2 ) / LEN
It's not difficult to notice that:
SquareOfRollingRMS[i] = RollingRMS[i-1] * LEN - a[i-1]^2 + a[i+LEN-1]^2
RollingRMS[i] = SqurefOfRollingRMS[i]^(1/2)
Doing it this way ,you are avoiding recaculating the overlap windows.
EDIT:
You can save some divide and multiply operation by moving LEN to the left side of the equations. This might speed up a lot as dividing is usually relatively slow.
LEN_by_SquareOfRollingRMS[0] = (a[0]^2 + a[1]^2 + a[2]^2)
LEN_by_SquareOfRollingRMS[i] = LEN_by_RollingRMS[i-1] - a[i-1]^2 + a[i+LEN-1]^2