mc2-scripts-and-benchmarks.git - Benchmarks and execution scripts for testing spatial isolation and SMT interference in MC^2

diff options

Diffstat (limited to 'all_pairs/source/h264_dec/copyright.txt')

0 files changed, 0 insertions, 0 deletions

210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344

// Copyright 2025 Joshua Bakita // Show or change the GPU core affinity for a CUDA process // taskset-like utility for NVIDIA GPUs #define _GNU_SOURCE // For program_invocation_name #include <argp.h> #include <dirent.h> #include <errno.h> #include <error.h> #include <fcntl.h> #include <stdbool.h> #include <stdint.h> #include <stdio.h> #include <stdlib.h> #include <string.h> #include <sys/mman.h> #include <unistd.h> #include <cuda.h> // To help with getting GPC info #include "libsmctrl.h" #define LINK_NAME "/memfd:libsmctrl" // TODO: Write automated tests: // - Change region of non-existent PID // - Change region of permission denied PID // - Change region of non-GPU PID // - Change GPC list // - Change TPC list // - Change TPC mask // - Start TPC mask // - Start TPC list // - Start GPC list // - Start with subargument containing - // - Query GPC list // - Query TPC list // - Query TPC mask // - Set GPC list w/ non-existant GPC // - Set TPC list w/ non-existant TPC // Private symbols from libsmctrl extern bool libsmctrl_is_mps_running(); extern uint128_t strtou128(const char *nptr, char **endptr, int base); const char *argp_program_bug_address = "<jbakita@cs.unc.edu>"; const char *argp_program_version = "nvtaskset 2025.06"; const char desc[] = "Show or change the GPU core affinity for a CUDA process\v" "Warning: When using GPC lists, this tool currently " "derives TPC to GPC mappings from the first NVIDIA GPU in " "the system (by PCI bus ID) device. To use the mappings " "for a different device, use the `libsmctrl_test_get_info` " "tool to get the bitmask of TPCs associated with each GPC, " "OR them, and then set that bitmask via this tool. Better " "multi-GPU support is intended for a future release.\n\n" "Inspired by the Linux taskset utility."; const char args_doc[] = "[mask | list] [pid | cmd [args...]]"; const struct argp_option opts[] = { {"gpc-list", 'g', NULL, 0, "Specify partition as a list of GPCs"}, {"tpc-list", 't', NULL, 0, "Specify partition as a list of TPCs"}, {"pid", 'p', NULL, 0, "Operate on an existing PID"}, {0} }; // Create a CUDA context and query the associated GPC to TPC mappings // Based off logic in libsmctrl_test_gpc_info void libsmctrl_get_gpc_info_ext_easy(uint32_t* num_gpcs, uint128_t** masks, int gpu_id) { int res; CUcontext ctx; char *old_order = NULL; // Tell CUDA to use PCI device id ordering (to match nvdebug) putenv((char*)"CUDA_DEVICE_ORDER=PCI_BUS_ID"); // Allow CUDA to see all devices (to better match nvdebug) if (getenv("CUDA_VISIBLE_DEVICES")) { if (!(old_order = strdup(getenv("CUDA_VISIBLE_DEVICES")))) error(1, errno, "Unable to allocate environment string"); unsetenv("CUDA_VISIBLE_DEVICES"); } // A CUDA context is required before reading the topology information if ((res = cuInit(0))) { const char* name; cuGetErrorName(res, &name); error(1, 0, "Unable to create a initialize CUDA, error %s\n", name); } if ((res = cuCtxCreate(&ctx, 0, gpu_id))) { const char* name; cuGetErrorName(res, &name); error(1, 0, "Unable to create a CUDA context, error %s\n", name); } // Pull topology information from libsmctrl if ((res = libsmctrl_get_gpc_info_ext(num_gpcs, masks, gpu_id)) != 0) { error(0, res, "libsmctrl_get_gpc_info() failed"); if (res == ENOENT) fprintf(stderr, "%s: Is the nvdebug kernel module loaded?\n", program_invocation_name); if (res == EIO) fprintf(stderr, "%s: Is the GPU powered on, i.e., is there an active context?\n", program_invocation_name); exit(1); } // Restore the environment (in case we exec() later) unsetenv("CUDA_DEVICE_ORDER"); if (old_order) { setenv("CUDA_VISIBLE_DEVICES", old_order, 1); free(old_order); } } int parse_list(bool use_gpcs, char* list, uint128_t *mask_out) { // We support the same ranges as taskset, e.g., X,Y,Z and X,Y-Z uint32_t num_xpcs = 0; // Either TPC or GPC count, i.e., "X"PC uint128_t* masks = NULL; // TODO: Allow specifying GPU ID, rather than assuming 0! if (use_gpcs) libsmctrl_get_gpc_info_ext_easy(&num_xpcs, &masks, 0); else libsmctrl_get_tpc_info_cuda(&num_xpcs, 0); uint128_t mask = 0; int range_start_xpc = -1; char* start = list; int len = strlen(list); // Convert comma-seperated GPC/TPC list into a mask for (int i = 0; i < len + 1; i++) { if (list[i] == ',' || list[i] == '\0') { list[i] = '\0'; int xpc = atoi(start); if (xpc > num_xpcs - 1) error(1, EINVAL, "%s is not a valid %s ID", start, use_gpcs ? "GPC" : "TPC"); // Handle ranges if (range_start_xpc != -1) { if (range_start_xpc >= xpc) error(1, EINVAL, "Malformed %s range", use_gpcs ? "GPC" : "TPC"); while (range_start_xpc <= xpc) { if (use_gpcs) mask |= masks[range_start_xpc]; else mask |= (uint128_t)1 << range_start_xpc; range_start_xpc++; } range_start_xpc = -1; } else { if (use_gpcs) mask |= masks[xpc]; else mask |= (uint128_t)1 << xpc; } start = list + i + 1; } // Range start if (list[i] == '-') { list[i] = '\0'; range_start_xpc = atoi(start); start = list + i + 1; } } *mask_out = mask; return 0; } // Always returns a valid string char* compose_list(uint128_t mask) { // List will always be shorter than every TPC, comma-seperated // 128 TPCs, with 10 1-char, 90 2-char, 28 3-char, 127 commas, and 1 null static char list[10 + 90*2 + 28*3 + 128]; char* tail = list; int last_enabled = -2; bool in_range; for (int i = 0; i < 128; i++) { bool enabled = (mask >> i) & 1; if (in_range) { if (enabled) { last_enabled = i; } else { tail += sprintf(tail, "%d,", last_enabled); in_range = false; } continue; } if (enabled) { if (last_enabled == i - 1) { in_range = true; tail += sprintf(tail, "-"); } else { tail += sprintf(tail, "%d", i); } last_enabled = i; } else { if (last_enabled == i - 1) { tail += sprintf(tail, ","); } } } // Strip trailing comma if (*(tail - 1) == ',') *(tail - 1) = '\0'; return list; } // Always returns a valid string // (Terminates the program on error) char* compose_gpc_list(uint128_t mask) { uint32_t num_gpcs = 0; uint128_t* masks = NULL; libsmctrl_get_gpc_info_ext_easy(&num_gpcs, &masks, 0); uint128_t gpc_mask = 0; // Try to find correspondence between a list of TPCs and GPCs


context:
space:
mode: