1 package net.bmahe.genetics4j.gpu.opencl.model; 2 3 import org.immutables.value.Value; 4 5 /** 6 * Represents kernel-specific execution characteristics and resource requirements for an OpenCL kernel on a specific device. 7 * 8 * <p>KernelInfo encapsulates the device-specific compilation and execution characteristics of an OpenCL kernel, 9 * providing essential information for optimal work group configuration and resource allocation in GPU-accelerated 10 * evolutionary algorithms. This information is determined at kernel compilation time and varies by device. 11 * 12 * <p>Key kernel characteristics include: 13 * <ul> 14 * <li><strong>Work group constraints</strong>: Maximum and preferred work group sizes for efficient execution</li> 15 * <li><strong>Memory usage</strong>: Local and private memory requirements per work-item</li> 16 * <li><strong>Performance optimization</strong>: Preferred work group size multiples for optimal resource utilization</li> 17 * <li><strong>Resource validation</strong>: Constraints for validating kernel launch parameters</li> 18 * </ul> 19 * 20 * <p>Kernel optimization considerations for evolutionary algorithms: 21 * <ul> 22 * <li><strong>Work group sizing</strong>: Configure launch parameters within device-specific limits</li> 23 * <li><strong>Memory allocation</strong>: Ensure sufficient local memory for parallel fitness evaluation</li> 24 * <li><strong>Performance tuning</strong>: Align work group sizes with preferred multiples</li> 25 * <li><strong>Resource planning</strong>: Account for per-work-item memory requirements</li> 26 * </ul> 27 * 28 * <p>Common usage patterns for kernel configuration: 29 * <pre>{@code 30 * // Query kernel information after compilation 31 * KernelInfo kernelInfo = kernelInfoReader.read(deviceId, kernel, "fitness_evaluation"); 32 * 33 * // Configure work group size within device limits 34 * long maxWorkGroupSize = Math.min(kernelInfo.workGroupSize(), device.maxWorkGroupSize()); 35 * 36 * // Optimize for preferred work group size multiple 37 * long preferredMultiple = kernelInfo.preferredWorkGroupSizeMultiple(); 38 * long optimalWorkGroupSize = (maxWorkGroupSize / preferredMultiple) * preferredMultiple; 39 * 40 * // Validate memory requirements for population size 41 * long populationSize = 1000; 42 * long totalLocalMem = kernelInfo.localMemSize() * optimalWorkGroupSize; 43 * long totalPrivateMem = kernelInfo.privateMemSize() * populationSize; 44 * 45 * // Configure kernel execution with validated parameters 46 * clEnqueueNDRangeKernel(commandQueue, kernel, 1, null, 47 * new long[]{populationSize}, new long[]{optimalWorkGroupSize}, 0, null, null); 48 * }</pre> 49 * 50 * <p>Performance optimization workflow: 51 * <ol> 52 * <li><strong>Kernel compilation</strong>: Compile kernel for target device</li> 53 * <li><strong>Information query</strong>: Read kernel-specific execution characteristics</li> 54 * <li><strong>Work group optimization</strong>: Calculate optimal work group size based on preferences</li> 55 * <li><strong>Memory validation</strong>: Ensure memory requirements fit within device limits</li> 56 * <li><strong>Launch configuration</strong>: Configure kernel execution with optimized parameters</li> 57 * </ol> 58 * 59 * <p>Memory management considerations: 60 * <ul> 61 * <li><strong>Local memory</strong>: Shared among work-items in the same work group</li> 62 * <li><strong>Private memory</strong>: Individual memory per work-item</li> 63 * <li><strong>Total allocation</strong>: Sum of all work-items' memory requirements</li> 64 * <li><strong>Device limits</strong>: Validate against device memory constraints</li> 65 * </ul> 66 * 67 * <p>Error handling and validation: 68 * <ul> 69 * <li><strong>Work group limits</strong>: Ensure launch parameters don't exceed kernel limits</li> 70 * <li><strong>Memory constraints</strong>: Validate total memory usage against device capabilities</li> 71 * <li><strong>Performance degradation</strong>: Monitor for suboptimal work group configurations</li> 72 * <li><strong>Resource conflicts</strong>: Handle multiple kernels competing for device resources</li> 73 * </ul> 74 * 75 * @see Device 76 * @see net.bmahe.genetics4j.gpu.opencl.KernelInfoReader 77 * @see net.bmahe.genetics4j.gpu.opencl.KernelInfoUtils 78 */ 79 @Value.Immutable 80 public interface KernelInfo { 81 82 /** 83 * Returns the name of the kernel function. 84 * 85 * @return the kernel function name as specified in the OpenCL program 86 */ 87 String name(); 88 89 /** 90 * Returns the maximum work group size that can be used when executing this kernel on the device. 91 * 92 * <p>This value represents the maximum number of work-items that can be in a work group when 93 * executing this specific kernel on the target device. It may be smaller than the device's 94 * general maximum work group size due to kernel-specific resource requirements. 95 * 96 * @return the maximum work group size for this kernel 97 */ 98 long workGroupSize(); 99 100 /** 101 * Returns the preferred work group size multiple for optimal kernel execution performance. 102 * 103 * <p>For optimal performance, the work group size should be a multiple of this value. 104 * This represents the native vector width or wavefront size of the device and helps 105 * achieve better resource utilization and memory coalescing. 106 * 107 * @return the preferred work group size multiple for performance optimization 108 */ 109 long preferredWorkGroupSizeMultiple(); 110 111 /** 112 * Returns the amount of local memory in bytes used by this kernel. 113 * 114 * <p>Local memory is shared among all work-items in a work group and includes both 115 * statically allocated local variables and dynamically allocated local memory passed 116 * as kernel arguments. This value is used to validate that the total local memory 117 * usage doesn't exceed the device's local memory capacity. 118 * 119 * @return the local memory usage in bytes per work group 120 */ 121 long localMemSize(); 122 123 /** 124 * Returns the minimum amount of private memory in bytes used by each work-item. 125 * 126 * <p>Private memory is individual to each work-item and includes local variables, 127 * function call stacks, and other per-work-item data. This value helps estimate 128 * the total memory footprint when launching kernels with large work group sizes. 129 * 130 * @return the private memory usage in bytes per work-item 131 */ 132 long privateMemSize(); 133 134 /** 135 * Creates a new builder for constructing KernelInfo instances. 136 * 137 * @return a new builder for creating kernel information objects 138 */ 139 static ImmutableKernelInfo.Builder builder() { 140 return ImmutableKernelInfo.builder(); 141 } 142 }