/// <summary>
///
/// </summary>
/// <param name="input"></param>
/// <returns></returns>
public static int SequentialAddressing(int[] input)
{
// Preconditions
if (input == null)
{
throw new ArgumentNullException("input");
}
// Declare variables to hold the current number of threads per block and number of blocks; re-used throughout the code below
int NumThreadsPerBlock = 0, NumBlocks = 0;
// Calculate the block and grid sizes needed for the first "level" of the reduction
CalculateBlockAndGridSizes(2, input.Length, out NumThreadsPerBlock, out NumBlocks);
Launcher.SetBlockSize(NumThreadsPerBlock);
Launcher.SetGridSize(NumBlocks);
// Perform the first "level" of the reduction
int[] OutputData = new int[NumBlocks];
Reduction.SequentialAddressingKernel(input, OutputData);
// If necessary, perform additional "levels" of reduction until we have only a single element (the result)
while (NumBlocks > CpuThreshold)
{
// Calculate the number of threads and blocks based on the current input size
CalculateBlockAndGridSizes(2, NumBlocks, out NumThreadsPerBlock, out NumBlocks);
Launcher.SetBlockSize(NumThreadsPerBlock);
Launcher.SetGridSize(NumBlocks);
// Replace the current "level's" input data with the output data from the previous "level"
input = OutputData;
// Create a new array to hold the output data for this "level"
OutputData = new int[NumBlocks];
// Call the reduction method to perform the current "level" of reduction
Reduction.SequentialAddressingKernel(input, OutputData);
}
// Return the final reduction value
return(OutputData[0]);
}
/// <summary>
///
/// </summary>
/// <param name="input"></param>
/// <returns></returns>
public static int FirstReductionFromGlobal(int[] input)
{
// Preconditions
if (input == null)
{
throw new ArgumentNullException("input");
}
//// Perform the first "level" of the reduction
//OutputData = new int[NumBlocks];
//Reduction.FirstReductionFromGlobalKernel(ClonedInputData, OutputData);
//// If necessary, perform additional "levels" of reduction until we have only a single element (the result)
//while (NumBlocks > CpuThreshold)
//{
// // TEMP: Calculate the result from the partial block sums (for testing purposes)
// int TempResult = 0;
// for (int i = 0; i < OutputData.Length; i++) { TempResult += OutputData[i]; }
// // Calculate the number of threads and blocks based on the current input size
// CalculateBlockAndGridSizes(3, NumBlocks, out NumThreadsPerBlock, out NumBlocks);
// Launcher.SetBlockSize(NumThreadsPerBlock);
// Launcher.SetGridSize(NumBlocks);
// // Replace the current "level's" input data with the output data from the previous "level"
// ClonedInputData = OutputData;
// // Create a new array to hold the output data for this "level"
// OutputData = new int[NumBlocks];
// // Call the reduction method to perform the current "level" of reduction
// Reduction.FirstReductionFromGlobalKernel(ClonedInputData, OutputData);
//}
//Watch.Stop();
//Console.Out.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", OutputData[0], Watch.Elapsed.TotalMilliseconds);
//Console.Out.Write("Test ");
//if (OutputData[0] == CpuReductionValue) { ConsoleWriteLineColored("passed!", ConsoleColor.Green); }
//else { ConsoleWriteLineColored("failed!", ConsoleColor.Red); }
//Console.Out.WriteLine();
//Watch.Reset();
// Declare variables to hold the current number of threads per block and number of blocks; re-used throughout the code below
int NumThreadsPerBlock = 0, NumBlocks = 0;
// Calculate the block and grid sizes needed for the first "level" of the reduction
CalculateBlockAndGridSizes(3, input.Length, out NumThreadsPerBlock, out NumBlocks);
Launcher.SetBlockSize(NumThreadsPerBlock);
Launcher.SetGridSize(NumBlocks);
// Perform the first "level" of the reduction
int[] OutputData = new int[NumBlocks];
Reduction.FirstReductionFromGlobalKernel(input, OutputData);
// If necessary, perform additional "levels" of reduction until we have only a single element (the result)
while (NumBlocks > CpuThreshold)
{
// Calculate the number of threads and blocks based on the current input size
CalculateBlockAndGridSizes(3, NumBlocks, out NumThreadsPerBlock, out NumBlocks);
Launcher.SetBlockSize(NumThreadsPerBlock);
Launcher.SetGridSize(NumBlocks);
// Replace the current "level's" input data with the output data from the previous "level"
input = OutputData;
// Create a new array to hold the output data for this "level"
OutputData = new int[NumBlocks];
// Call the reduction method to perform the current "level" of reduction
Reduction.FirstReductionFromGlobalKernel(input, OutputData);
}
// Return the final reduction value
return(OutputData[0]);
}
private const int NumElements = 1 << 23;// 1 << 24; // 1 << 22 = 4194304 // 1 << 24 = 16777216 // Must be power-of-two sizes
static void Main(string[] args)
{
// Create input and output data arrays
int[] InputData = new int[NumElements];
// Print 'header'
Console.WriteLine("Performing shared-memory reduction tests...");
Console.WriteLine("----------------------------------------------------------------------");
Console.Write("Generating random test data ({0} {1} elements)...", NumElements, InputData.GetType().GetElementType().Name);
// Fill the input data with random values; these values must fall between zero (0) and the maximum number, which if multiplied by the array length, would still fit in an int32 (signed int) value
// This is to make our results easier to validate, since we don't have to deal with possible overflow behavior
Random rand = new Random();
const int MaxValue = Int32.MaxValue / NumElements;
for (int i = 0; i < InputData.Length; i++)
{
InputData[i] = rand.Next(MaxValue);
}
Console.WriteLine("done.");
Console.WriteLine();
// Create the stopwatch we'll use to time how long each reduction takes
Stopwatch Watch = new Stopwatch();
// Compute the reduction value on the CPU first so that we can compare it to the GPU-based results
// TODO: Perform the reduction 2 or 3 times here to get an accurate timing result
// TODO: Create a version of this project which uses PLINQ / TPL for comparison
Console.WriteLine("Computing CPU-based result for comparison...");
Watch.Start();
int CpuReductionValue = 0;
for (int i = 0; i < InputData.Length; i++)
{
CpuReductionValue += InputData[i];
}
Watch.Stop();
Console.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", CpuReductionValue, Watch.Elapsed.TotalMilliseconds);
Console.WriteLine();
Console.WriteLine("----------------------------------------------------------------------");
Watch.Reset();
#region reduce0 (Interleaved access with modulo operator)
// Start the reduction (and the timer)
Console.WriteLine("Testing reduce0 (Interleaved access with modulo operator)...");
Watch.Start();
// Call the reduction method, which will iterate the reduction kernel until the entire array is reduced
int InterleavedModuloResult = Reduction.InterleavedModulo(InputData);
Watch.Stop();
Console.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", InterleavedModuloResult, Watch.Elapsed.TotalMilliseconds);
Console.Write("Test ");
if (InterleavedModuloResult == CpuReductionValue)
{
ConsoleWriteLineColored("passed!", ConsoleColor.Green);
}
else
{
ConsoleWriteLineColored("failed!", ConsoleColor.Red);
}
Console.WriteLine();
Watch.Reset();
#endregion
#region reduce1 (Interleaved contiguous access)
// Start the reduction (and the timer)
Console.WriteLine("Testing reduce1 (Interleaved contiguous access)...");
Watch.Start();
// Call the reduction method, which will iterate the reduction kernel until the entire array is reduced
int InterleavedContiguousResult = Reduction.InterleavedContiguous(InputData);
Watch.Stop();
Console.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", InterleavedContiguousResult, Watch.Elapsed.TotalMilliseconds);
Console.Write("Test ");
if (InterleavedContiguousResult == CpuReductionValue)
{
ConsoleWriteLineColored("passed!", ConsoleColor.Green);
}
else
{
ConsoleWriteLineColored("failed!", ConsoleColor.Red);
}
Console.WriteLine();
Watch.Reset();
#endregion
#region reduce2 (Sequential addressing)
// Start the reduction (and the timer)
Console.WriteLine("Testing reduce2 (Sequential addressing)...");
Watch.Start();
// Call the reduction method, which will iterate the reduction kernel until the entire array is reduced
int SequentialAddressingResult = Reduction.SequentialAddressing(InputData);
Watch.Stop();
Console.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", SequentialAddressingResult, Watch.Elapsed.TotalMilliseconds);
Console.Write("Test ");
if (SequentialAddressingResult == CpuReductionValue)
{
ConsoleWriteLineColored("passed!", ConsoleColor.Green);
}
else
{
ConsoleWriteLineColored("failed!", ConsoleColor.Red);
}
Console.WriteLine();
Watch.Reset();
#endregion
#region reduce3 (Sequential addressing with first reduction from global)
// Start the reduction (and the timer)
Console.WriteLine("Testing reduce3 (Sequential addressing with reduction from global)...");
Watch.Start();
// Call the reduction method, which will iterate the reduction kernel until the entire array is reduced
int FirstReductionFromGlobalResult = Reduction.FirstReductionFromGlobal(InputData);
Watch.Stop();
Console.WriteLine("done. (Value = {0}, Time = {1:F02} ms)", FirstReductionFromGlobalResult, Watch.Elapsed.TotalMilliseconds);
Console.Write("Test ");
if (FirstReductionFromGlobalResult == CpuReductionValue)
{
ConsoleWriteLineColored("passed!", ConsoleColor.Green);
}
else
{
ConsoleWriteLineColored("failed!", ConsoleColor.Red);
}
Console.WriteLine();
Watch.Reset();
#endregion
// Print the exit message
Console.WriteLine("Press any key to exit.");
Console.ReadKey();
Environment.Exit(0);
}