├── README.md ├── css └── index.css ├── img └── fork_me_right_gray_6d6d6d.png ├── index.html └── js ├── calculator.coffee ├── calculator.js ├── calculator.js.map └── index.js /README.md: -------------------------------------------------------------------------------- 1 | # Online CUDA Occupancy Calculator 2 | 3 | https://xmartlabs.github.io/cuda-calculator/ 4 | 5 | Port of the [CUDA Occupancy Calculator spreadsheet](https://docs.nvidia.com/cuda/cuda-occupancy-calculator/CUDA_Occupancy_Calculator.xls). 6 | 7 | ## Changes from upstream 8 | 9 | * Extended until Compute Capabilities 8.6 10 | * Improved the UI. 11 | 12 | ## Compilation 13 | 14 | ```shell 15 | coffee -c js/calculator.coffee 16 | ``` 17 | 18 | ## Credits 19 | 20 | * [Xmartlabs](https://xmartlabs.github.io/cuda-calculator/) 21 | * [Karthikeyan](https://github.com/lxkarthi/cuda-calculator/) - Improved cuda-calculator online version. 22 | * [Mihai Maruseac](https://github.com/mihaimaruseac/cuda-calculator/) - Thanks for the fork. 23 | * [Aliaksei](https://github.com/roadhump/cuda-calculator/) - Original Author deleted it. :( 24 | -------------------------------------------------------------------------------- /css/index.css: -------------------------------------------------------------------------------- 1 | caption { 2 | caption-side: top; 3 | } 4 | 5 | .table td { 6 | text-align: right; 7 | } 8 | 9 | .c3-axis-x-label, .c3-axis-y-label, .c3-xgrid-line { 10 | font-size: 1rem; 11 | } 12 | 13 | #fork-me { 14 | position: absolute; 15 | top: 0; 16 | right: 0; 17 | border: 0; 18 | } 19 | -------------------------------------------------------------------------------- /img/fork_me_right_gray_6d6d6d.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xmartlabs/cuda-calculator/d60116f321cda7cfd414c260d3133e48b4feceea/img/fork_me_right_gray_6d6d6d.png -------------------------------------------------------------------------------- /index.html: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | CUDA Occupancy Calculator 7 | 8 | 9 | 11 | 14 | 15 | 16 | 17 | 18 | 19 | 20 |

21 |

CUDA Occupancy Calculator

22 | 23 | 70 | 71 |

72 | 73 | 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | 83 | 84 | 85 | 86 | 87 | 88 | 89 | 90 |

GPU Occupancy Data is displayed here and in the graphs
Active Threads per Multiprocessor
Active Warps per Multiprocessor
Active Thread Blocks per Multiprocessor
Occupancy of each Multiprocessor

91 | 92 | 93 | 94 | 95 | 96 | 97 | 98 | 99 | 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | 108 | 109 | 110 | 111 | 112 | 113 | 114 | 115 | 116 | 117 | 118 | 119 | 120 | 121 | 122 | 123 | 124 | 125 | 126 | 127 | 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | 150 |

Physical Limits for GPU Compute Capability
Version
Threads per Warp
Warps per Multiprocessor
Threads per Multiprocessor
Thread Blocks per Multiprocessor
Total # of 32-bit registers per Multiprocessor
Register allocation unit size
Register allocation granularity
Max registers per Block
Max registers per thread
Shared Memory per Multiprocessor (bytes)
Shared Memory Allocation unit size
Warp allocation granularity (for register allocation)
Max thread block size

151 | 152 | 153 | 154 | 155 | 156 | 157 | 158 | 159 | 160 | 161 | 162 | 163 | 164 | 165 | 166 |

Allocation Per Thread Block
Warps
Registers
Shared Memory

167 |

168 | Note: CUDA Runtime uses bytes of Shared Memory per Thread Block. 169 |

170 | 171 | 172 | 173 | 174 | 175 | 176 | 177 | 178 | 179 | 180 | 181 | 182 | 183 | 184 | 185 |

Maximum Thread Blocks Per Multiprocessor
Limited by Max Warps / Blocks per Multiprocessor
Limited by Registers per Multiprocessor
Limited by Shared Memory per Multiprocessor

186 | 187 |

188 |

Impact of Varying Block Size

189 |

190 | 191 | 192 |

193 |

Impact of Varying Register Count Per Thread

194 |

195 | 196 | 197 |

198 |

Impact of Varying Shared Memory Usage Per Block

199 |

200 | 201 |

202 |

203 | 204 | 205 | Fork me on GitHub

206 | 207 | 208 | 211 | 212 | 215 | 216 | 217 | 218 | 219 | 220 | 221 | -------------------------------------------------------------------------------- /js/calculator.coffee: -------------------------------------------------------------------------------- 1 | mainConfig = 2 | '2.0': 3 | version: '2.0' 4 | threadsPerWarp: 32 5 | warpsPerMultiprocessor: 48 6 | threadsPerMultiprocessor: 1536 7 | threadBlocksPerMultiprocessor: 8 8 | sharedMemoryPerMultiprocessor: 49152 9 | 10 | registerFileSize: 32768 11 | registerAllocationUnitSize: 64 12 | 13 | allocationGranularity: 'warp' 14 | maxRegistersPerThread: 63 15 | maxRegistersPerBlock: 32768 16 | sharedMemoryAllocationUnitSize: 128 17 | warpAllocationGranularity: 2 18 | maxThreadBlockSize: 1024 19 | 20 | '2.1': 21 | version: '2.1' 22 | threadsPerWarp: 32 23 | warpsPerMultiprocessor: 48 24 | threadsPerMultiprocessor: 1536 25 | threadBlocksPerMultiprocessor: 8 26 | sharedMemoryPerMultiprocessor: 49152 27 | 28 | registerFileSize: 32768 29 | registerAllocationUnitSize: 64 30 | 31 | allocationGranularity: 'warp' 32 | maxRegistersPerThread: 63 33 | maxRegistersPerBlock: 32768 34 | sharedMemoryAllocationUnitSize: 128 35 | warpAllocationGranularity: 2 36 | maxThreadBlockSize: 1024 37 | 38 | '3.0': 39 | version: '3.0' 40 | threadsPerWarp: 32 41 | warpsPerMultiprocessor: 64 42 | threadsPerMultiprocessor: 2048 43 | threadBlocksPerMultiprocessor: 16 44 | sharedMemoryPerMultiprocessor: 49152 45 | 46 | registerFileSize: 65536 47 | registerAllocationUnitSize: 256 48 | 49 | allocationGranularity: 'warp' 50 | maxRegistersPerThread: 63 51 | maxRegistersPerBlock: 65536 52 | sharedMemoryAllocationUnitSize: 256 53 | warpAllocationGranularity: 4 54 | maxThreadBlockSize: 1024 55 | 56 | '3.2': 57 | version: '3.2' 58 | threadsPerWarp: 32 59 | warpsPerMultiprocessor: 64 60 | threadsPerMultiprocessor: 2048 61 | threadBlocksPerMultiprocessor: 16 62 | sharedMemoryPerMultiprocessor: 49152 63 | 64 | registerFileSize: 65536 65 | registerAllocationUnitSize: 256 66 | 67 | allocationGranularity: 'warp' 68 | maxRegistersPerThread: 255 69 | maxRegistersPerBlock: 65536 70 | sharedMemoryAllocationUnitSize: 256 71 | warpAllocationGranularity: 4 72 | maxThreadBlockSize: 1024 73 | 74 | '3.5': 75 | version: '3.5' 76 | threadsPerWarp: 32 77 | warpsPerMultiprocessor: 64 78 | threadsPerMultiprocessor: 2048 79 | threadBlocksPerMultiprocessor: 16 80 | sharedMemoryPerMultiprocessor: 49152 81 | 82 | registerFileSize: 65536 83 | registerAllocationUnitSize: 256 84 | 85 | allocationGranularity: 'warp' 86 | maxRegistersPerThread: 255 87 | maxRegistersPerBlock: 65536 88 | sharedMemoryAllocationUnitSize: 256 89 | warpAllocationGranularity: 4 90 | maxThreadBlockSize: 1024 91 | 92 | '3.7': 93 | version: '3.7' 94 | threadsPerWarp: 32 95 | warpsPerMultiprocessor: 64 96 | threadsPerMultiprocessor: 2048 97 | threadBlocksPerMultiprocessor: 16 98 | sharedMemoryPerMultiprocessor: 114688 99 | 100 | registerFileSize: 131072 101 | registerAllocationUnitSize: 256 102 | 103 | allocationGranularity: 'warp' 104 | maxRegistersPerThread: 255 105 | maxRegistersPerBlock: 65536 106 | sharedMemoryAllocationUnitSize: 256 107 | warpAllocationGranularity: 4 108 | maxThreadBlockSize: 1024 109 | 110 | '5.0': 111 | version: '5.0' 112 | threadsPerWarp: 32 113 | warpsPerMultiprocessor: 64 114 | threadsPerMultiprocessor: 2048 115 | threadBlocksPerMultiprocessor: 32 116 | sharedMemoryPerMultiprocessor: 65536 117 | 118 | registerFileSize: 65536 119 | registerAllocationUnitSize: 256 120 | 121 | allocationGranularity: 'warp' 122 | maxRegistersPerThread: 255 123 | maxRegistersPerBlock: 65536 124 | sharedMemoryAllocationUnitSize: 256 125 | warpAllocationGranularity: 4 126 | maxThreadBlockSize: 1024 127 | 128 | '5.2': 129 | version: '5.2' 130 | threadsPerWarp: 32 131 | warpsPerMultiprocessor: 64 132 | threadsPerMultiprocessor: 2048 133 | threadBlocksPerMultiprocessor: 32 134 | sharedMemoryPerMultiprocessor: 98304 135 | 136 | registerFileSize: 65536 137 | registerAllocationUnitSize: 256 138 | 139 | allocationGranularity: 'warp' 140 | maxRegistersPerThread: 255 141 | maxRegistersPerBlock: 32768 142 | sharedMemoryAllocationUnitSize: 256 143 | warpAllocationGranularity: 4 144 | maxThreadBlockSize: 1024 145 | 146 | '5.3': 147 | version: '5.3' 148 | threadsPerWarp: 32 149 | warpsPerMultiprocessor: 64 150 | threadsPerMultiprocessor: 2048 151 | threadBlocksPerMultiprocessor: 32 152 | sharedMemoryPerMultiprocessor: 65536 153 | 154 | registerFileSize: 65536 155 | registerAllocationUnitSize: 256 156 | 157 | allocationGranularity: 'warp' 158 | maxRegistersPerThread: 255 159 | maxRegistersPerBlock: 32768 160 | sharedMemoryAllocationUnitSize: 256 161 | warpAllocationGranularity: 4 162 | maxThreadBlockSize: 1024 163 | 164 | '6.0': 165 | version: '6.0' 166 | threadsPerWarp: 32 167 | warpsPerMultiprocessor: 64 168 | threadsPerMultiprocessor: 2048 169 | threadBlocksPerMultiprocessor: 32 170 | sharedMemoryPerMultiprocessor: 65536 171 | 172 | registerFileSize: 65536 173 | registerAllocationUnitSize: 256 174 | 175 | allocationGranularity: 'warp' 176 | maxRegistersPerThread: 255 177 | maxRegistersPerBlock: 65536 178 | sharedMemoryAllocationUnitSize: 256 179 | warpAllocationGranularity: 2 180 | maxThreadBlockSize: 1024 181 | 182 | '6.1': 183 | version: '6.1' 184 | threadsPerWarp: 32 185 | warpsPerMultiprocessor: 64 186 | threadsPerMultiprocessor: 2048 187 | threadBlocksPerMultiprocessor: 32 188 | sharedMemoryPerMultiprocessor: 98304 189 | 190 | registerFileSize: 65536 191 | registerAllocationUnitSize: 256 192 | 193 | allocationGranularity: 'warp' 194 | maxRegistersPerThread: 255 195 | maxRegistersPerBlock: 65536 196 | sharedMemoryAllocationUnitSize: 256 197 | warpAllocationGranularity: 4 198 | maxThreadBlockSize: 1024 199 | 200 | '6.2': 201 | version: '6.2' 202 | threadsPerWarp: 32 203 | warpsPerMultiprocessor: 64 204 | threadsPerMultiprocessor: 2048 205 | threadBlocksPerMultiprocessor: 32 206 | sharedMemoryPerMultiprocessor: 65536 207 | 208 | registerFileSize: 65536 209 | registerAllocationUnitSize: 256 210 | 211 | allocationGranularity: 'warp' 212 | maxRegistersPerThread: 255 213 | maxRegistersPerBlock: 65536 214 | sharedMemoryAllocationUnitSize: 256 215 | warpAllocationGranularity: 4 216 | maxThreadBlockSize: 1024 217 | 218 | '7.0': 219 | version: '7.0' 220 | threadsPerWarp: 32 221 | warpsPerMultiprocessor: 64 222 | threadsPerMultiprocessor: 2048 223 | threadBlocksPerMultiprocessor: 32 224 | sharedMemoryPerMultiprocessor: 98304 225 | 226 | registerFileSize: 65536 227 | registerAllocationUnitSize: 256 228 | 229 | allocationGranularity: 'warp' 230 | maxRegistersPerThread: 255 231 | maxRegistersPerBlock: 65536 232 | sharedMemoryAllocationUnitSize: 256 233 | warpAllocationGranularity: 4 234 | maxThreadBlockSize: 1024 235 | 236 | '7.5': 237 | version: '7.5' 238 | threadsPerWarp: 32 239 | warpsPerMultiprocessor: 32 240 | threadsPerMultiprocessor: 1024 241 | threadBlocksPerMultiprocessor: 16 242 | sharedMemoryPerMultiprocessor: 65536 243 | 244 | registerFileSize: 65536 245 | registerAllocationUnitSize: 256 246 | 247 | allocationGranularity: 'warp' 248 | maxRegistersPerThread: 255 249 | maxRegistersPerBlock: 65536 250 | sharedMemoryAllocationUnitSize: 256 251 | warpAllocationGranularity: 4 252 | maxThreadBlockSize: 1024 253 | 254 | '8.0': 255 | version: '8.0' 256 | threadsPerWarp: 32 257 | warpsPerMultiprocessor: 64 258 | threadsPerMultiprocessor: 2048 259 | threadBlocksPerMultiprocessor: 32 260 | sharedMemoryPerMultiprocessor: 167936 261 | 262 | registerFileSize: 65536 263 | registerAllocationUnitSize: 256 264 | 265 | allocationGranularity: 'warp' 266 | maxRegistersPerThread: 255 267 | maxRegistersPerBlock: 65536 268 | sharedMemoryAllocationUnitSize: 128 269 | warpAllocationGranularity: 4 270 | maxThreadBlockSize: 1024 271 | 272 | '8.6': 273 | version: '8.6' 274 | threadsPerWarp: 32 275 | warpsPerMultiprocessor: 48 276 | threadsPerMultiprocessor: 1536 277 | threadBlocksPerMultiprocessor: 16 278 | sharedMemoryPerMultiprocessor: 102400 279 | 280 | registerFileSize: 65536 281 | registerAllocationUnitSize: 256 282 | 283 | allocationGranularity: 'warp' 284 | maxRegistersPerThread: 255 285 | maxRegistersPerBlock: 65536 286 | sharedMemoryAllocationUnitSize: 128 287 | warpAllocationGranularity: 4 288 | maxThreadBlockSize: 1024 289 | 290 | # Starting with Compute Capability 8.x, CUDA runtime uses 1KB of Shared Memory 291 | cudaRuntimeUsedSharedMemory = 292 | '11.0': 1024 293 | '11.1': 1024 294 | 295 | ceil = (a, b) -> Math.ceil(a / b) * b 296 | 297 | floor = (a, b) -> Math.floor(a / b) * b 298 | 299 | 300 | window.calculateOccupancy = (input) -> 301 | config = mainConfig[input.version] 302 | 303 | # number of warps per block 304 | blockWarps = () -> 305 | Math.ceil(input.threadsPerBlock / config.threadsPerWarp) 306 | 307 | # number of registers per warp 308 | registersPerWarp = () -> 309 | ceil(input.registersPerThread * config.threadsPerWarp, config.registerAllocationUnitSize) 310 | 311 | # number of registers per block 312 | blockRegisters = () -> 313 | registersPerWarp() * blockWarps() 314 | 315 | # maximum warps per SM when limited by registers 316 | warpsPerMultiprocessorLimitedByRegisters = () -> 317 | floor(config.maxRegistersPerBlock / registersPerWarp(), config.warpAllocationGranularity) 318 | 319 | # starting with Compute Capability 8.x, the CUDA runtime consumes 1KB of shared memory 320 | # the amount might change depending on the CUDA runtime version in the future 321 | blockCudaRuntimeSharedMemory = () -> 322 | if Number.parseFloat(input.version) >= 8 323 | cudaRuntimeUsedSharedMemory[input.cudaVersion] 324 | else 325 | 0 326 | 327 | # shared memory per thread block 328 | blockSharedMemory = () -> 329 | ceil( 330 | Number.parseInt(input.sharedMemoryPerBlock) + blockCudaRuntimeSharedMemory(), 331 | config.sharedMemoryAllocationUnitSize 332 | ) 333 | 334 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor = () -> 335 | Math.min( 336 | config.threadBlocksPerMultiprocessor, Math.floor(config.warpsPerMultiprocessor / blockWarps()) 337 | ) 338 | 339 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor = () -> 340 | if input.registersPerThread > config.maxRegistersPerThread 341 | 0 342 | else if input.registersPerThread > 0 343 | Math.floor(warpsPerMultiprocessorLimitedByRegisters() / blockWarps()) * 344 | Math.floor(config.registerFileSize / config.maxRegistersPerBlock) 345 | else 346 | config.threadBlocksPerMultiprocessor 347 | 348 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor = () -> 349 | if input.sharedMemoryPerBlock > 0 350 | Math.floor(config.sharedMemoryPerMultiprocessor / blockSharedMemory()) 351 | else 352 | config.threadBlocksPerMultiprocessor 353 | 354 | activeThreadsPerMultiprocessor = () -> 355 | input.threadsPerBlock * activeThreadBlocksPerMultiprocessor() 356 | 357 | activeWarpsPerMultiprocessor = () -> 358 | activeThreadBlocksPerMultiprocessor() * blockWarps() 359 | 360 | activeThreadBlocksPerMultiprocessor = () -> 361 | Math.min( 362 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor(), 363 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor(), 364 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor() 365 | ) 366 | 367 | occupancyOfMultiprocessor = () -> 368 | activeWarpsPerMultiprocessor() / config.warpsPerMultiprocessor 369 | 370 | output = 371 | activeThreadsPerMultiprocessor: activeThreadsPerMultiprocessor() 372 | activeWarpsPerMultiprocessor: activeWarpsPerMultiprocessor() 373 | activeThreadBlocksPerMultiprocessor: activeThreadBlocksPerMultiprocessor() 374 | occupancyOfMultiprocessor: occupancyOfMultiprocessor() 375 | 376 | blockWarps: blockWarps() 377 | blockSharedMemory: blockSharedMemory() 378 | blockCudaRuntimeSharedMemory: blockCudaRuntimeSharedMemory() 379 | blockRegisters: blockRegisters() 380 | 381 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor: 382 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor() 383 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor: 384 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor() 385 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor: 386 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor() 387 | 388 | return Object.assign(output, config) 389 | 390 | window.computeGraphsValues = (input) -> 391 | config = mainConfig[input.version] 392 | 393 | graphWarpOccupancyOfThreadsPerBlock = () -> 394 | current = 395 | key: input.threadsPerBlock 396 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 397 | 398 | inp = Object.assign({}, input) # Shallow copy. 399 | r = [] 400 | for threadsPerBlock in [32..config.maxThreadBlockSize] by 32 401 | inp.threadsPerBlock = threadsPerBlock 402 | 403 | r.push({ 404 | key: threadsPerBlock 405 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 406 | }) 407 | 408 | return { 409 | xLabel: "Threads per block" 410 | data: r 411 | current: current 412 | } 413 | 414 | graphWarpOccupancyOfRegistersPerThread = () -> 415 | current = 416 | key: input.registersPerThread 417 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 418 | 419 | inp = Object.assign({}, input) # Shallow copy. 420 | r = [] 421 | for registersPerThread in [0..config.maxRegistersPerThread] 422 | inp.registersPerThread = registersPerThread 423 | 424 | r.push({ 425 | key: registersPerThread 426 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 427 | }) 428 | 429 | return { 430 | xLabel: "Registers per thread" 431 | data: r 432 | current: current 433 | } 434 | 435 | graphWarpOccupancyOfSharedMemoryPerBlock = () -> 436 | current = 437 | key: input.sharedMemoryPerBlock 438 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 439 | 440 | inp = Object.assign({}, input) # Shallow copy. 441 | r = [] 442 | for sharedMemoryPerBlock in [0..config.sharedMemoryPerMultiprocessor] by 512 443 | inp.sharedMemoryPerBlock = sharedMemoryPerBlock 444 | 445 | r.push({ 446 | key: sharedMemoryPerBlock 447 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 448 | }) 449 | 450 | return { 451 | xLabel: "Shared memory per block" 452 | data: r 453 | current: current 454 | } 455 | 456 | return { 457 | graphWarpOccupancyOfThreadsPerBlock: graphWarpOccupancyOfThreadsPerBlock() 458 | graphWarpOccupancyOfRegistersPerThread: graphWarpOccupancyOfRegistersPerThread() 459 | graphWarpOccupancyOfSharedMemoryPerBlock: graphWarpOccupancyOfSharedMemoryPerBlock() 460 | } 461 | -------------------------------------------------------------------------------- /js/calculator.js: -------------------------------------------------------------------------------- 1 | // Generated by CoffeeScript 1.12.7 2 | (function() { 3 | var ceil, cudaRuntimeUsedSharedMemory, floor, mainConfig; 4 | 5 | mainConfig = { 6 | '2.0': { 7 | version: '2.0', 8 | threadsPerWarp: 32, 9 | warpsPerMultiprocessor: 48, 10 | threadsPerMultiprocessor: 1536, 11 | threadBlocksPerMultiprocessor: 8, 12 | sharedMemoryPerMultiprocessor: 49152, 13 | registerFileSize: 32768, 14 | registerAllocationUnitSize: 64, 15 | allocationGranularity: 'warp', 16 | maxRegistersPerThread: 63, 17 | maxRegistersPerBlock: 32768, 18 | sharedMemoryAllocationUnitSize: 128, 19 | warpAllocationGranularity: 2, 20 | maxThreadBlockSize: 1024 21 | }, 22 | '2.1': { 23 | version: '2.1', 24 | threadsPerWarp: 32, 25 | warpsPerMultiprocessor: 48, 26 | threadsPerMultiprocessor: 1536, 27 | threadBlocksPerMultiprocessor: 8, 28 | sharedMemoryPerMultiprocessor: 49152, 29 | registerFileSize: 32768, 30 | registerAllocationUnitSize: 64, 31 | allocationGranularity: 'warp', 32 | maxRegistersPerThread: 63, 33 | maxRegistersPerBlock: 32768, 34 | sharedMemoryAllocationUnitSize: 128, 35 | warpAllocationGranularity: 2, 36 | maxThreadBlockSize: 1024 37 | }, 38 | '3.0': { 39 | version: '3.0', 40 | threadsPerWarp: 32, 41 | warpsPerMultiprocessor: 64, 42 | threadsPerMultiprocessor: 2048, 43 | threadBlocksPerMultiprocessor: 16, 44 | sharedMemoryPerMultiprocessor: 49152, 45 | registerFileSize: 65536, 46 | registerAllocationUnitSize: 256, 47 | allocationGranularity: 'warp', 48 | maxRegistersPerThread: 63, 49 | maxRegistersPerBlock: 65536, 50 | sharedMemoryAllocationUnitSize: 256, 51 | warpAllocationGranularity: 4, 52 | maxThreadBlockSize: 1024 53 | }, 54 | '3.2': { 55 | version: '3.2', 56 | threadsPerWarp: 32, 57 | warpsPerMultiprocessor: 64, 58 | threadsPerMultiprocessor: 2048, 59 | threadBlocksPerMultiprocessor: 16, 60 | sharedMemoryPerMultiprocessor: 49152, 61 | registerFileSize: 65536, 62 | registerAllocationUnitSize: 256, 63 | allocationGranularity: 'warp', 64 | maxRegistersPerThread: 255, 65 | maxRegistersPerBlock: 65536, 66 | sharedMemoryAllocationUnitSize: 256, 67 | warpAllocationGranularity: 4, 68 | maxThreadBlockSize: 1024 69 | }, 70 | '3.5': { 71 | version: '3.5', 72 | threadsPerWarp: 32, 73 | warpsPerMultiprocessor: 64, 74 | threadsPerMultiprocessor: 2048, 75 | threadBlocksPerMultiprocessor: 16, 76 | sharedMemoryPerMultiprocessor: 49152, 77 | registerFileSize: 65536, 78 | registerAllocationUnitSize: 256, 79 | allocationGranularity: 'warp', 80 | maxRegistersPerThread: 255, 81 | maxRegistersPerBlock: 65536, 82 | sharedMemoryAllocationUnitSize: 256, 83 | warpAllocationGranularity: 4, 84 | maxThreadBlockSize: 1024 85 | }, 86 | '3.7': { 87 | version: '3.7', 88 | threadsPerWarp: 32, 89 | warpsPerMultiprocessor: 64, 90 | threadsPerMultiprocessor: 2048, 91 | threadBlocksPerMultiprocessor: 16, 92 | sharedMemoryPerMultiprocessor: 114688, 93 | registerFileSize: 131072, 94 | registerAllocationUnitSize: 256, 95 | allocationGranularity: 'warp', 96 | maxRegistersPerThread: 255, 97 | maxRegistersPerBlock: 65536, 98 | sharedMemoryAllocationUnitSize: 256, 99 | warpAllocationGranularity: 4, 100 | maxThreadBlockSize: 1024 101 | }, 102 | '5.0': { 103 | version: '5.0', 104 | threadsPerWarp: 32, 105 | warpsPerMultiprocessor: 64, 106 | threadsPerMultiprocessor: 2048, 107 | threadBlocksPerMultiprocessor: 32, 108 | sharedMemoryPerMultiprocessor: 65536, 109 | registerFileSize: 65536, 110 | registerAllocationUnitSize: 256, 111 | allocationGranularity: 'warp', 112 | maxRegistersPerThread: 255, 113 | maxRegistersPerBlock: 65536, 114 | sharedMemoryAllocationUnitSize: 256, 115 | warpAllocationGranularity: 4, 116 | maxThreadBlockSize: 1024 117 | }, 118 | '5.2': { 119 | version: '5.2', 120 | threadsPerWarp: 32, 121 | warpsPerMultiprocessor: 64, 122 | threadsPerMultiprocessor: 2048, 123 | threadBlocksPerMultiprocessor: 32, 124 | sharedMemoryPerMultiprocessor: 98304, 125 | registerFileSize: 65536, 126 | registerAllocationUnitSize: 256, 127 | allocationGranularity: 'warp', 128 | maxRegistersPerThread: 255, 129 | maxRegistersPerBlock: 32768, 130 | sharedMemoryAllocationUnitSize: 256, 131 | warpAllocationGranularity: 4, 132 | maxThreadBlockSize: 1024 133 | }, 134 | '5.3': { 135 | version: '5.3', 136 | threadsPerWarp: 32, 137 | warpsPerMultiprocessor: 64, 138 | threadsPerMultiprocessor: 2048, 139 | threadBlocksPerMultiprocessor: 32, 140 | sharedMemoryPerMultiprocessor: 65536, 141 | registerFileSize: 65536, 142 | registerAllocationUnitSize: 256, 143 | allocationGranularity: 'warp', 144 | maxRegistersPerThread: 255, 145 | maxRegistersPerBlock: 32768, 146 | sharedMemoryAllocationUnitSize: 256, 147 | warpAllocationGranularity: 4, 148 | maxThreadBlockSize: 1024 149 | }, 150 | '6.0': { 151 | version: '6.0', 152 | threadsPerWarp: 32, 153 | warpsPerMultiprocessor: 64, 154 | threadsPerMultiprocessor: 2048, 155 | threadBlocksPerMultiprocessor: 32, 156 | sharedMemoryPerMultiprocessor: 65536, 157 | registerFileSize: 65536, 158 | registerAllocationUnitSize: 256, 159 | allocationGranularity: 'warp', 160 | maxRegistersPerThread: 255, 161 | maxRegistersPerBlock: 65536, 162 | sharedMemoryAllocationUnitSize: 256, 163 | warpAllocationGranularity: 2, 164 | maxThreadBlockSize: 1024 165 | }, 166 | '6.1': { 167 | version: '6.1', 168 | threadsPerWarp: 32, 169 | warpsPerMultiprocessor: 64, 170 | threadsPerMultiprocessor: 2048, 171 | threadBlocksPerMultiprocessor: 32, 172 | sharedMemoryPerMultiprocessor: 98304, 173 | registerFileSize: 65536, 174 | registerAllocationUnitSize: 256, 175 | allocationGranularity: 'warp', 176 | maxRegistersPerThread: 255, 177 | maxRegistersPerBlock: 65536, 178 | sharedMemoryAllocationUnitSize: 256, 179 | warpAllocationGranularity: 4, 180 | maxThreadBlockSize: 1024 181 | }, 182 | '6.2': { 183 | version: '6.2', 184 | threadsPerWarp: 32, 185 | warpsPerMultiprocessor: 64, 186 | threadsPerMultiprocessor: 2048, 187 | threadBlocksPerMultiprocessor: 32, 188 | sharedMemoryPerMultiprocessor: 65536, 189 | registerFileSize: 65536, 190 | registerAllocationUnitSize: 256, 191 | allocationGranularity: 'warp', 192 | maxRegistersPerThread: 255, 193 | maxRegistersPerBlock: 65536, 194 | sharedMemoryAllocationUnitSize: 256, 195 | warpAllocationGranularity: 4, 196 | maxThreadBlockSize: 1024 197 | }, 198 | '7.0': { 199 | version: '7.0', 200 | threadsPerWarp: 32, 201 | warpsPerMultiprocessor: 64, 202 | threadsPerMultiprocessor: 2048, 203 | threadBlocksPerMultiprocessor: 32, 204 | sharedMemoryPerMultiprocessor: 98304, 205 | registerFileSize: 65536, 206 | registerAllocationUnitSize: 256, 207 | allocationGranularity: 'warp', 208 | maxRegistersPerThread: 255, 209 | maxRegistersPerBlock: 65536, 210 | sharedMemoryAllocationUnitSize: 256, 211 | warpAllocationGranularity: 4, 212 | maxThreadBlockSize: 1024 213 | }, 214 | '7.5': { 215 | version: '7.5', 216 | threadsPerWarp: 32, 217 | warpsPerMultiprocessor: 32, 218 | threadsPerMultiprocessor: 1024, 219 | threadBlocksPerMultiprocessor: 16, 220 | sharedMemoryPerMultiprocessor: 65536, 221 | registerFileSize: 65536, 222 | registerAllocationUnitSize: 256, 223 | allocationGranularity: 'warp', 224 | maxRegistersPerThread: 255, 225 | maxRegistersPerBlock: 65536, 226 | sharedMemoryAllocationUnitSize: 256, 227 | warpAllocationGranularity: 4, 228 | maxThreadBlockSize: 1024 229 | }, 230 | '8.0': { 231 | version: '8.0', 232 | threadsPerWarp: 32, 233 | warpsPerMultiprocessor: 64, 234 | threadsPerMultiprocessor: 2048, 235 | threadBlocksPerMultiprocessor: 32, 236 | sharedMemoryPerMultiprocessor: 167936, 237 | registerFileSize: 65536, 238 | registerAllocationUnitSize: 256, 239 | allocationGranularity: 'warp', 240 | maxRegistersPerThread: 255, 241 | maxRegistersPerBlock: 65536, 242 | sharedMemoryAllocationUnitSize: 128, 243 | warpAllocationGranularity: 4, 244 | maxThreadBlockSize: 1024 245 | }, 246 | '8.6': { 247 | version: '8.6', 248 | threadsPerWarp: 32, 249 | warpsPerMultiprocessor: 48, 250 | threadsPerMultiprocessor: 1536, 251 | threadBlocksPerMultiprocessor: 16, 252 | sharedMemoryPerMultiprocessor: 102400, 253 | registerFileSize: 65536, 254 | registerAllocationUnitSize: 256, 255 | allocationGranularity: 'warp', 256 | maxRegistersPerThread: 255, 257 | maxRegistersPerBlock: 65536, 258 | sharedMemoryAllocationUnitSize: 128, 259 | warpAllocationGranularity: 4, 260 | maxThreadBlockSize: 1024 261 | } 262 | }; 263 | 264 | cudaRuntimeUsedSharedMemory = { 265 | '11.0': 1024, 266 | '11.1': 1024 267 | }; 268 | 269 | ceil = function(a, b) { 270 | return Math.ceil(a / b) * b; 271 | }; 272 | 273 | floor = function(a, b) { 274 | return Math.floor(a / b) * b; 275 | }; 276 | 277 | window.calculateOccupancy = function(input) { 278 | var activeThreadBlocksPerMultiprocessor, activeThreadsPerMultiprocessor, activeWarpsPerMultiprocessor, blockCudaRuntimeSharedMemory, blockRegisters, blockSharedMemory, blockWarps, config, occupancyOfMultiprocessor, output, registersPerWarp, threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor, threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor, threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor, warpsPerMultiprocessorLimitedByRegisters; 279 | config = mainConfig[input.version]; 280 | blockWarps = function() { 281 | return Math.ceil(input.threadsPerBlock / config.threadsPerWarp); 282 | }; 283 | registersPerWarp = function() { 284 | return ceil(input.registersPerThread * config.threadsPerWarp, config.registerAllocationUnitSize); 285 | }; 286 | blockRegisters = function() { 287 | return registersPerWarp() * blockWarps(); 288 | }; 289 | warpsPerMultiprocessorLimitedByRegisters = function() { 290 | return floor(config.maxRegistersPerBlock / registersPerWarp(), config.warpAllocationGranularity); 291 | }; 292 | blockCudaRuntimeSharedMemory = function() { 293 | if (Number.parseFloat(input.version) >= 8) { 294 | return cudaRuntimeUsedSharedMemory[input.cudaVersion]; 295 | } else { 296 | return 0; 297 | } 298 | }; 299 | blockSharedMemory = function() { 300 | return ceil(Number.parseInt(input.sharedMemoryPerBlock) + blockCudaRuntimeSharedMemory(), config.sharedMemoryAllocationUnitSize); 301 | }; 302 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor = function() { 303 | return Math.min(config.threadBlocksPerMultiprocessor, Math.floor(config.warpsPerMultiprocessor / blockWarps())); 304 | }; 305 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor = function() { 306 | if (input.registersPerThread > config.maxRegistersPerThread) { 307 | return 0; 308 | } else if (input.registersPerThread > 0) { 309 | return Math.floor(warpsPerMultiprocessorLimitedByRegisters() / blockWarps()) * Math.floor(config.registerFileSize / config.maxRegistersPerBlock); 310 | } else { 311 | return config.threadBlocksPerMultiprocessor; 312 | } 313 | }; 314 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor = function() { 315 | if (input.sharedMemoryPerBlock > 0) { 316 | return Math.floor(config.sharedMemoryPerMultiprocessor / blockSharedMemory()); 317 | } else { 318 | return config.threadBlocksPerMultiprocessor; 319 | } 320 | }; 321 | activeThreadsPerMultiprocessor = function() { 322 | return input.threadsPerBlock * activeThreadBlocksPerMultiprocessor(); 323 | }; 324 | activeWarpsPerMultiprocessor = function() { 325 | return activeThreadBlocksPerMultiprocessor() * blockWarps(); 326 | }; 327 | activeThreadBlocksPerMultiprocessor = function() { 328 | return Math.min(threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor(), threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor(), threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor()); 329 | }; 330 | occupancyOfMultiprocessor = function() { 331 | return activeWarpsPerMultiprocessor() / config.warpsPerMultiprocessor; 332 | }; 333 | output = { 334 | activeThreadsPerMultiprocessor: activeThreadsPerMultiprocessor(), 335 | activeWarpsPerMultiprocessor: activeWarpsPerMultiprocessor(), 336 | activeThreadBlocksPerMultiprocessor: activeThreadBlocksPerMultiprocessor(), 337 | occupancyOfMultiprocessor: occupancyOfMultiprocessor(), 338 | blockWarps: blockWarps(), 339 | blockSharedMemory: blockSharedMemory(), 340 | blockCudaRuntimeSharedMemory: blockCudaRuntimeSharedMemory(), 341 | blockRegisters: blockRegisters(), 342 | threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor: threadBlocksPerMultiprocessorLimitedByWarpsOrBlocksPerMultiprocessor(), 343 | threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor: threadBlocksPerMultiprocessorLimitedByRegistersPerMultiprocessor(), 344 | threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor: threadBlocksPerMultiprocessorLimitedBySharedMemoryPerMultiprocessor() 345 | }; 346 | return Object.assign(output, config); 347 | }; 348 | 349 | window.computeGraphsValues = function(input) { 350 | var config, graphWarpOccupancyOfRegistersPerThread, graphWarpOccupancyOfSharedMemoryPerBlock, graphWarpOccupancyOfThreadsPerBlock; 351 | config = mainConfig[input.version]; 352 | graphWarpOccupancyOfThreadsPerBlock = function() { 353 | var current, i, inp, r, ref, threadsPerBlock; 354 | current = { 355 | key: input.threadsPerBlock, 356 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 357 | }; 358 | inp = Object.assign({}, input); 359 | r = []; 360 | for (threadsPerBlock = i = 32, ref = config.maxThreadBlockSize; i <= ref; threadsPerBlock = i += 32) { 361 | inp.threadsPerBlock = threadsPerBlock; 362 | r.push({ 363 | key: threadsPerBlock, 364 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 365 | }); 366 | } 367 | return { 368 | xLabel: "Threads per block", 369 | data: r, 370 | current: current 371 | }; 372 | }; 373 | graphWarpOccupancyOfRegistersPerThread = function() { 374 | var current, i, inp, r, ref, registersPerThread; 375 | current = { 376 | key: input.registersPerThread, 377 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 378 | }; 379 | inp = Object.assign({}, input); 380 | r = []; 381 | for (registersPerThread = i = 0, ref = config.maxRegistersPerThread; 0 <= ref ? i <= ref : i >= ref; registersPerThread = 0 <= ref ? ++i : --i) { 382 | inp.registersPerThread = registersPerThread; 383 | r.push({ 384 | key: registersPerThread, 385 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 386 | }); 387 | } 388 | return { 389 | xLabel: "Registers per thread", 390 | data: r, 391 | current: current 392 | }; 393 | }; 394 | graphWarpOccupancyOfSharedMemoryPerBlock = function() { 395 | var current, i, inp, r, ref, sharedMemoryPerBlock; 396 | current = { 397 | key: input.sharedMemoryPerBlock, 398 | value: window.calculateOccupancy(input).activeWarpsPerMultiprocessor 399 | }; 400 | inp = Object.assign({}, input); 401 | r = []; 402 | for (sharedMemoryPerBlock = i = 0, ref = config.sharedMemoryPerMultiprocessor; i <= ref; sharedMemoryPerBlock = i += 512) { 403 | inp.sharedMemoryPerBlock = sharedMemoryPerBlock; 404 | r.push({ 405 | key: sharedMemoryPerBlock, 406 | value: window.calculateOccupancy(inp).activeWarpsPerMultiprocessor 407 | }); 408 | } 409 | return { 410 | xLabel: "Shared memory per block", 411 | data: r, 412 | current: current 413 | }; 414 | }; 415 | return { 416 | graphWarpOccupancyOfThreadsPerBlock: graphWarpOccupancyOfThreadsPerBlock(), 417 | graphWarpOccupancyOfRegistersPerThread: graphWarpOccupancyOfRegistersPerThread(), 418 | graphWarpOccupancyOfSharedMemoryPerBlock: graphWarpOccupancyOfSharedMemoryPerBlock() 419 | }; 420 | }; 421 | 422 | }).call(this); 423 | -------------------------------------------------------------------------------- /js/calculator.js.map: -------------------------------------------------------------------------------- 1 | { 2 | "version": 3, 3 | "file": "calculator.js", 4 | "sourceRoot": "", 5 | "sources": [ 6 | "calculator.coffee" 7 | ], 8 | "names": [], 9 | "mappings": ";AAAA;AAAA,MAAA;;EAAA,UAAA,GACE;IAAA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,GAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,IAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,OAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,GAdpB;KADF;IAiBA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,GAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,IAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,OAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,GAdpB;KAlBF;IAkCA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,OAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,GAdpB;KAnCF;IAmDA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,OAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,GAdpB;KApDF;IAoEA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,EAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,EAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KArEF;IAqFA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,CAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,EAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,EAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAtFF;IAsGA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,EAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAvGF;IAuHA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAxHF;IAwIA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,MAL/B;MAOA,gBAAA,EAAkB,MAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAzIF;IAyJA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA1JF;IA0KA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA3KF;IA2LA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA5LF;IA4MA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA7MF;IA6NA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA9NF;IA8OA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KA/OF;IA+PA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAhQF;IAgRA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,KAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAjRF;IAiSA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,MAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAlSF;IAkTA,KAAA,EACE;MAAA,OAAA,EAAS,KAAT;MACA,cAAA,EAAgB,EADhB;MAEA,sBAAA,EAAwB,EAFxB;MAGA,wBAAA,EAA0B,IAH1B;MAIA,6BAAA,EAA+B,EAJ/B;MAKA,6BAAA,EAA+B,MAL/B;MAOA,gBAAA,EAAkB,KAPlB;MAQA,0BAAA,EAA4B,GAR5B;MAUA,qBAAA,EAAuB,MAVvB;MAWA,qBAAA,EAAuB,GAXvB;MAYA,8BAAA,EAAgC,GAZhC;MAaA,yBAAA,EAA2B,CAb3B;MAcA,kBAAA,EAAoB,IAdpB;KAnTF;;;EAoUF,IAAA,GAAO,SAAC,CAAD,EAAI,CAAJ;WAAU,IAAI,CAAC,IAAL,CAAU,CAAA,GAAI,CAAd,CAAA,GAAmB;EAA7B;;EAEP,KAAA,GAAQ,SAAC,CAAD,EAAI,CAAJ;WAAU,IAAI,CAAC,KAAL,CAAW,CAAA,GAAI,CAAf,CAAA,GAAoB;EAA9B;;EAGR,MAAM,CAAC,kBAAP,GAA4B,SAAC,KAAD;AAC1B,QAAA;IAAA,MAAA,GAAS,UAAW,CAAA,KAAK,CAAC,OAAN;IAEpB,UAAA,GAAa,SAAA;aACX,IAAI,CAAC,IAAL,CAAU,KAAK,CAAC,eAAN,GAAwB,MAAM,CAAC,cAAzC;IADW;IAGb,cAAA,GAAiB,SAAA;MACf,IAAG,MAAM,CAAC,qBAAP,KAAgC,OAAnC;eACE,IAAA,CAAK,IAAA,CAAK,UAAA,CAAA,CAAL,EAAmB,MAAM,CAAC,yBAA1B,CAAA,GAAuD,KAAK,CAAC,kBAA7D,GAAkF,MAAM,CAAC,cAA9F,EAA8G,MAAM,CAAC,0BAArH,EADF;OAAA,MAAA;eAIE,IAAA,CAAK,KAAK,CAAC,kBAAN,GAA2B,MAAM,CAAC,cAAvC,EAAuD,MAAM,CAAC,0BAA9D,CAAA,GAA4F,UAAA,CAAA,EAJ9F;;IADe;IAOjB,uBAAA,GAA0B,SAAA;MACxB,IAAG,MAAM,CAAC,qBAAP,KAAgC,OAAnC;eACE,MAAM,CAAC,iBADT;OAAA,MAAA;eAIE,KAAA,CAAM,MAAM,CAAC,gBAAP,GAA0B,IAAA,CAAK,KAAK,CAAC,kBAAN,GAA2B,MAAM,CAAC,cAAvC,EAC9B,MAAM,CAAC,0BADuB,CAAhC,EACsC,MAAM,CAAC,yBAD7C,CAAA,GAEE,IAAA,CAAK,KAAK,CAAC,kBAAN,GAA2B,MAAM,CAAC,cAAvC,EAAuD,MAAM,CAAC,0BAA9D,EANJ;;IADwB;IAS1B,iBAAA,GAAoB,SAAA;aAClB,IAAA,CAAK,KAAK,CAAC,oBAAX,EAAiC,MAAM,CAAC,8BAAxC;IADkB;IAGpB,oEAAA,GAAuE,SAAA;aACrE,IAAI,CAAC,GAAL,CAAS,MAAM,CAAC,6BAAhB,EAA+C,IAAI,CAAC,KAAL,CAAW,MAAM,CAAC,sBAAP,GAAgC,UAAA,CAAA,CAA3C,CAA/C;IADqE;IAGvE,gEAAA,GAAmE,SAAA;MACjE,IAAG,KAAK,CAAC,kBAAN,GAA2B,MAAM,CAAC,qBAArC;eACE,EADF;OAAA,MAEK,IAAG,KAAK,CAAC,kBAAN,GAA2B,CAA9B;eACH,IAAI,CAAC,KAAL,CAAW,uBAAA,CAAA,CAAA,GAA4B,cAAA,CAAA,CAAvC,EADG;OAAA,MAAA;eAGH,MAAM,CAAC,8BAHJ;;IAH4D;IAQnE,mEAAA,GAAsE,SAAA;MACpE,IAAG,KAAK,CAAC,oBAAN,GAA6B,CAAhC;eACE,IAAI,CAAC,KAAL,CAAW,MAAM,CAAC,6BAAP,GAAuC,iBAAA,CAAA,CAAlD,EADF;OAAA,MAAA;eAGE,MAAM,CAAC,8BAHT;;IADoE;IAMtE,8BAAA,GAAiC,SAAA;aAC/B,KAAK,CAAC,eAAN,GAAwB,mCAAA,CAAA;IADO;IAGjC,4BAAA,GAA+B,SAAA;aAC7B,mCAAA,CAAA,CAAA,GAAwC,UAAA,CAAA;IADX;IAG/B,mCAAA,GAAsC,SAAA;aACpC,IAAI,CAAC,GAAL,CACE,oEAAA,CAAA,CADF,EAEE,gEAAA,CAAA,CAFF,EAGE,mEAAA,CAAA,CAHF;IADoC;IAOtC,yBAAA,GAA4B,SAAA;aAC1B,4BAAA,CAAA,CAAA,GAAiC,MAAM,CAAC;IADd;IAG5B,MAAA,GACE;MAAA,8BAAA,EAAgC,8BAAA,CAAA,CAAhC;MACA,4BAAA,EAA8B,4BAAA,CAAA,CAD9B;MAEA,mCAAA,EAAqC,mCAAA,CAAA,CAFrC;MAGA,yBAAA,EAA2B,yBAAA,CAAA,CAH3B;MAKA,UAAA,EAAY,UAAA,CAAA,CALZ;MAMA,iBAAA,EAAmB,iBAAA,CAAA,CANnB;MAOA,cAAA,EAAgB,cAAA,CAAA,CAPhB;MASA,oEAAA,EAAsE,oEAAA,CAAA,CATtE;MAUA,gEAAA,EAAkE,gEAAA,CAAA,CAVlE;MAWA,mEAAA,EAAqE,mEAAA,CAAA,CAXrE;;AAaF,WAAO,MAAM,CAAC,MAAP,CAAc,MAAd,EAAsB,MAAtB;EAxEmB;;EA0E5B,MAAM,CAAC,mBAAP,GAA6B,SAAC,KAAD;AAC3B,QAAA;IAAA,MAAA,GAAS,UAAW,CAAA,KAAK,CAAC,OAAN;IAEpB,mCAAA,GAAsC,SAAA;AACpC,UAAA;MAAA,OAAA,GACE;QAAA,GAAA,EAAK,KAAK,CAAC,eAAX;QACA,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,KAA1B,CAAgC,CAAC,4BADxC;;MAGF,GAAA,GAAM,MAAM,CAAC,MAAP,CAAc,EAAd,EAAkB,KAAlB;MACN,CAAA,GAAI;AACJ,WAAuB,8FAAvB;QACE,GAAG,CAAC,eAAJ,GAAsB;QAEtB,CAAC,CAAC,IAAF,CAAO;UACL,GAAA,EAAK,eADA;UAEL,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,GAA1B,CAA8B,CAAC,4BAFjC;SAAP;AAHF;AAQA,aAAO;QACL,MAAA,EAAQ,mBADH;QAEL,IAAA,EAAM,CAFD;QAGL,OAAA,EAAS,OAHJ;;IAf6B;IAqBtC,sCAAA,GAAyC,SAAA;AACvC,UAAA;MAAA,OAAA,GACE;QAAA,GAAA,EAAK,KAAK,CAAC,kBAAX;QACA,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,KAA1B,CAAgC,CAAC,4BADxC;;MAGF,GAAA,GAAM,MAAM,CAAC,MAAP,CAAc,EAAd,EAAkB,KAAlB;MACN,CAAA,GAAI;AACJ,WAA0B,yIAA1B;QACE,GAAG,CAAC,kBAAJ,GAAyB;QAEzB,CAAC,CAAC,IAAF,CAAO;UACL,GAAA,EAAK,kBADA;UAEL,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,GAA1B,CAA8B,CAAC,4BAFjC;SAAP;AAHF;AAQA,aAAO;QACL,MAAA,EAAQ,sBADH;QAEL,IAAA,EAAM,CAFD;QAGL,OAAA,EAAS,OAHJ;;IAfgC;IAqBzC,wCAAA,GAA2C,SAAA;AACzC,UAAA;MAAA,OAAA,GACE;QAAA,GAAA,EAAK,KAAK,CAAC,oBAAX;QACA,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,KAA1B,CAAgC,CAAC,4BADxC;;MAGF,GAAA,GAAM,MAAM,CAAC,MAAP,CAAc,EAAd,EAAkB,KAAlB;MACN,CAAA,GAAI;AACJ,WAA4B,mHAA5B;QACE,GAAG,CAAC,oBAAJ,GAA2B;QAE3B,CAAC,CAAC,IAAF,CAAO;UACL,GAAA,EAAK,oBADA;UAEL,KAAA,EAAO,MAAM,CAAC,kBAAP,CAA0B,GAA1B,CAA8B,CAAC,4BAFjC;SAAP;AAHF;AAQA,aAAO;QACL,MAAA,EAAQ,yBADH;QAEL,IAAA,EAAM,CAFD;QAGL,OAAA,EAAS,OAHJ;;IAfkC;AAqB3C,WAAO;MACL,mCAAA,EAAqC,mCAAA,CAAA,CADhC;MAEL,sCAAA,EAAwC,sCAAA,CAAA,CAFnC;MAGL,wCAAA,EAA0C,wCAAA,CAAA,CAHrC;;EAlEoB;AApZ7B" 10 | } -------------------------------------------------------------------------------- /js/index.js: -------------------------------------------------------------------------------- 1 | /*! 2 | * Serialize all form data into an array of key/value pairs 3 | * (c) 2020 Chris Ferdinandi, MIT License, https://gomakethings.com 4 | * @param {Node} form The form to serialize 5 | * @return {Array} The serialized form data 6 | */ 7 | function serializeArray(form) { 8 | const arr = []; 9 | Array.prototype.slice.call(form.elements).forEach(function (field) { 10 | if (!field.name || field.disabled || ['file', 'reset', 'submit', 'button'].indexOf(field.type) > -1) { 11 | return; 12 | } 13 | if (field.type === 'select-multiple') { 14 | Array.prototype.slice.call(field.options).forEach(function (option) { 15 | if (!option.selected) { 16 | return; 17 | } 18 | arr.push({ 19 | name: field.name, 20 | value: option.value 21 | }); 22 | }); 23 | return; 24 | } 25 | if (['checkbox', 'radio'].indexOf(field.type) > -1 && !field.checked) { 26 | return; 27 | } 28 | arr.push({ 29 | name: field.name, 30 | value: field.value 31 | }); 32 | }); 33 | return arr; 34 | } 35 | 36 | function docReady(fn) { 37 | if (document.readyState === "complete" || document.readyState === "interactive") { 38 | setTimeout(fn, 1); // Call on next available tick. 39 | } else { 40 | document.addEventListener("DOMContentLoaded", fn); 41 | } 42 | } 43 | 44 | function plot(x, y, currentX, currentY, xLabel, targetElement) { 45 | const chart = c3.generate({ 46 | bindto: targetElement, 47 | data: { 48 | xs: { 49 | " ": "x", 50 | "Current": "x current", 51 | }, 52 | columns: [ 53 | ["x"].concat(x), 54 | [" "].concat(y), 55 | 56 | ["x current", currentX], 57 | ["Current", currentY], 58 | ], 59 | }, 60 | axis: { 61 | x: { 62 | label: xLabel, 63 | min: 0, 64 | padding: { 65 | left: 0, 66 | right: 0, 67 | }, 68 | }, 69 | y: { 70 | label: { 71 | position: "outer-middle", 72 | text: "# warps", 73 | }, 74 | min: 0, 75 | padding: { 76 | bottom: 0, 77 | top: 0, 78 | }, 79 | }, 80 | }, 81 | grid: { 82 | y: { 83 | show: true, 84 | }, 85 | }, 86 | legend: { 87 | show: false, 88 | }, 89 | padding: { 90 | right: 20, 91 | } 92 | }); 93 | 94 | // To have no padding in axes but preventing the points from being dropped. 95 | d3.select(chart.element).select("." + c3.chart.internal.fn.CLASS.chart).attr("clip-path", null); 96 | } 97 | 98 | function onSubmit(e) { 99 | e.preventDefault(); 100 | 101 | const formElement = e.target; 102 | 103 | const formData = Object.fromEntries(serializeArray(formElement).map(n => [n.name, n.value])); 104 | 105 | const occupancyCalculationOutput = calculateOccupancy(formData); 106 | const graphsValues = computeGraphsValues(formData); 107 | 108 | document.getElementById("output").removeAttribute("hidden"); 109 | 110 | Object.entries(occupancyCalculationOutput) 111 | .forEach(([k, v]) => document.getElementById(k).innerText = v.toString()); 112 | 113 | Object.entries(graphsValues).forEach(([k, v]) => 114 | plot(v.data.map(d => d.key), v.data.map(d => d.value), v.current.key, v.current.value, v.xLabel, 115 | document.getElementById(k))); 116 | 117 | // show or hide the alert for Shared Memory used by Cuda Runtime 118 | const $alert = document.querySelector("#alertCudaRuntimeSharedMemory"); 119 | if (Number.parseFloat(occupancyCalculationOutput.version) >= 8.0) { 120 | $alert.removeAttribute("hidden"); 121 | } else { 122 | $alert.setAttribute("hidden", ""); 123 | } 124 | } 125 | 126 | function main() { 127 | document.getElementsByTagName("form")[0].onsubmit = onSubmit; 128 | 129 | // add event listener for changes on Compute Capability selector 130 | document.querySelector("#ccVersion").addEventListener('change', (event) => { 131 | // selected compute capability version 132 | const cc = Number.parseFloat(event.target.value); 133 | 134 | // get the div element with the selector for cuda version 135 | const $cudaVersionBlock = document.querySelector("#cudaVersion").closest("div"); 136 | 137 | // if CC >= 8.x, show the CUDA runtime version selector, else hide it 138 | if (cc >= 8) { 139 | $cudaVersionBlock.removeAttribute("hidden"); 140 | } else { 141 | $cudaVersionBlock.setAttribute("hidden", ""); 142 | } 143 | }); 144 | 145 | // set the default Compute Capability (latest version) 146 | const $ccSelect = document.querySelector("#ccVersion"); 147 | $ccSelect.selectedIndex = $ccSelect.options.length - 1; 148 | $ccSelect.dispatchEvent(new Event('change')); 149 | } 150 | 151 | docReady(main); 152 | --------------------------------------------------------------------------------