StockSharp Algo.Gpu Library

Overview

StockSharp.Algo.Gpu is a specialized library that provides GPU-accelerated calculations for technical indicators using the ILGPU framework. It enables massive parallel processing of market data indicators on CUDA, OpenCL, and CPU accelerators, offering significant performance improvements for batch calculations and strategy optimization scenarios.

Features

• GPU Acceleration – leverages ILGPU for cross-platform GPU computing with support for CUDA, OpenCL, and CPU backends
• Batch Processing – efficiently processes multiple data series and parameter combinations in a single GPU pass
• Optimized Data Structures – specialized GpuCandle and GpuIndicatorResult structs designed for GPU memory layout
• Technical Indicators – currently includes GPU-accelerated Simple Moving Average (SMA), Average Directional Index (ADX), and Envelope with extensible architecture for additional indicators
• Automatic Device Selection – intelligent accelerator selection based on device capabilities and memory size

Getting Started

Prerequisites

• .NET SDK 6.0 or later
• GPU with CUDA support, OpenCL drivers, or modern CPU for fallback processing
• Visual Studio 2022 or any compatible IDE

Basic Usage

The typical workflow involves creating an accelerator, setting up input data, and running batch calculations:

using StockSharp.Algo.Gpu;
using StockSharp.Algo.Gpu.Indicators;
using StockSharp.Algo.Indicators;

// Create best available accelerator (CUDA > OpenCL > CPU)
var (context, accelerator) = GpuAcceleratorFactory.CreateBestAccelerator();

// Convert candles to GPU format
var gpuCandles = candles.Select(c => new GpuCandle(
    c.OpenTime, c.OpenPrice, c.HighPrice, c.LowPrice, c.ClosePrice, c.TotalVolume
)).ToArray();

// Define SMA parameters to test
var smaParams = new GpuSmaParams[]
{
    new(10, (byte)Level1Fields.ClosePrice),
    new(20, (byte)Level1Fields.ClosePrice),
    new(50, (byte)Level1Fields.ClosePrice)
};

// Calculate SMA for multiple parameter sets
using var calculator = new GpuSmaCalculator(context, accelerator);
IGpuIndicatorResult[][][] gpuResults = calculator.Calculate(new[] { gpuCandles }, smaParams);

// Convert GPU results to indicator values
var smaIndicator = new SimpleMovingAverage { Length = 20 };
foreach (var seriesResult in gpuResults[0])
{
    foreach (IGpuIndicatorResult gpuResult in seriesResult)
    {
        if (gpuResult.GetIsFormed())
        {
            var indicatorValue = gpuResult.ToValue(smaIndicator);
            Console.WriteLine($"SMA: {indicatorValue.ToDecimal()} at {indicatorValue.Time}");
        }
    }
}

// Cleanup
accelerator.Dispose();
context.Dispose();

Architecture

Core Components

• GpuAcceleratorFactory – factory for creating and selecting optimal ILGPU accelerators
• GpuIndicatorCalculatorBase – base class for all GPU indicator calculators
• GpuCandle – GPU-optimized candle data structure with float precision for performance
• IGpuIndicatorResult – interface for GPU indicator calculation results with ToValue conversion
• GpuIndicatorResult – standard GPU result structure implementing IGpuIndicatorResult
• IGpuIndicatorParams – interface for GPU indicator parameter structures
• IGpuIndicatorResult – interface for GPU calculation results with ToValue() conversion method

Performance Considerations

• GPU calculations use single-precision floating-point (float) for optimal performance
• Batch processing is most efficient when processing multiple series or parameter combinations simultaneously
• Memory transfers between CPU and GPU are minimized through batched operations
• Automatic device selection prioritizes CUDA over OpenCL over CPU for best performance

Use Cases

• Strategy Optimization – testing multiple indicator parameters across large datasets
• Historical Analysis – processing years of market data with various technical indicators