Is it possible to terminate CUDA kernel from host? [duplicate] - cuda

Question

I am working with CUDA and I am trying to stop my kernels work (i.e. terminate all running threads) after a certain if block is being hit. How can I do that? I am really stuck in here.

Answer 1

The CUDA execution model doesn't allow for inter-block communication by design. That can potentially make this sort of kernel abort on condition operation difficult to achieve reliably without resorting to the assert or trap type approaches which can potentially result in context destruction and loss of data which isn't what you probably want.
If your kernel design involves a small number of blocks with "resident" threads, then the only approach is some sort of atomic spinlock, which is hard to get to work reliably, and which will greatly degrade memory controller performance and achievable bandwidth.
If, on the other hand, your kernel design has rather large grids with a lot of blocks, and your main goal is to stop blocks which are not yet scheduled from running, then you could try something like this:
#include <iostream>
#include <vector>
__device__ unsigned int found_idx;
__global__ void setkernel(unsigned int *indata)
{
indata[115949] = 0xdeadbeef;
indata[119086] = 0xdeadbeef;
indata[60534] = 0xdeadbeef;
indata[37072] = 0xdeadbeef;
indata[163107] = 0xdeadbeef;
}
__global__ void searchkernel(unsigned int *indata, unsigned int *outdata)
{
if (found_idx > 0) {
return;
} else if (threadIdx.x == 0) {
outdata[blockIdx.x] = blockIdx.x;
};
unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x;
if (indata[tid] == 0xdeadbeef) {
unsigned int oldval = atomicCAS(&found_idx, 0, 1+tid);
}
}
int main()
{
const unsigned int N = 1 << 19;
unsigned int* in_data;
cudaMalloc((void **)&in_data, sizeof(unsigned int) * size_t(N));
cudaMemset(in_data, 0, sizeof(unsigned int) * size_t(N));
setkernel<<<1,1>>>(in_data);
cudaDeviceSynchronize();
unsigned int block_size = 1024;
unsigned int grid_size = N / block_size;
unsigned int* out_data;
cudaMalloc((void **)&out_data, sizeof(unsigned int) * size_t(grid_size));
cudaMemset(out_data, 0xf0, sizeof(unsigned int) * size_t(grid_size));
const unsigned int zero = 0;
cudaMemcpyToSymbol(found_idx, &zero, sizeof(unsigned int));
searchkernel<<<grid_size, block_size>>>(in_data, out_data);
std::vector<unsigned int> output(grid_size);
cudaMemcpy(&output[0], out_data, sizeof(unsigned int) * size_t(grid_size), cudaMemcpyDeviceToHost);
cudaDeviceReset();
std::cout << "The following blocks did not run" << std::endl;
for(int i=0, j=0; i<grid_size; i++) {
if (output[i] == 0xf0f0f0f0) {
std::cout << " " << i;
if (j++ == 20) {
std::cout << std::endl;
j = 0;
}
}
}
std::cout << std::endl;
return 0;
}
Here I have a simple kernel which is searching for a magic word in a large array. To get the early exit behaviour, I use a single global word, which is set atomically by those threads which "win" or trigger the termination condition. Every new block checks the state of this global word, and if it is set, they return without doing any work.
If I compile and run this on a moderate sized Kepler device:
$ nvcc -arch=sm_30 -o blocking blocking.cu
$ ./blocking
The following blocks did not run
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
84 85 86 87 88 89 90 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
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
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
336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
504 505 506 507 508 509 510 511
you can see that a large number of blocks in the grid saw the change in the global word and early terminated without running the search code. This might be the best you can do without a severely invasive spinlock approach which will greatly harm performance.

Answer 2

I assume you want to stop a running kernel (not a single thread).
The simplest approach (and the one that I suggest) is to set up a global memory flag which is been tested by the kernel.
You can set the flag using cudaMemcpy() (or without if using unified memory).
Like the following:
if (gm_flag) {
__threadfence(); // ensure store issued before trap
asm("trap;"); // kill kernel with error
}
ams("trap;") will stop all running thread
Note that since cuda 2.0 you can use assert() to terminate a kernel!
A different approach could be the following (I haven't tried the code!)
__device__ bool go(int val){
return true;
}
__global__ void stopme(bool* flag, int* val, int size){
int idx= blockIdx.x *blockDim.x + threadIdx.x;
if(idx < size){
bool canContinue = true;
while(canContinue && (flag[0])){
printf("HELLO from %i\n",idx);
if(!(*flag)){
return;
}
else{
//do some computation
val[idx]++;
val[idx]%=100;
}
canContinue = go(val[idx]);
}
}
}
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
int main(void)
{
int size = 128;
int* h_val = (int*)malloc(sizeof(int)*size);
bool * h_flag = new bool;
*h_flag=true;
bool* d_flag;
cudaMalloc(&d_flag,sizeof(bool));
cudaMemcpy(d_flag,h_flag,1,cudaMemcpyHostToDevice);
int* d_val;
cudaMalloc(&d_val,sizeof(int)*size );
for(int i=0;i<size;i++){
h_val[i] = i;
}
cudaMemcpy(d_val,h_val,size,cudaMemcpyHostToDevice);
int BSIZE=32;
int nblocks =size/BSIZE;
printf("%i,%i",nblocks,BSIZE);
stopme<<<nblocks,BSIZE>>>(d_flag,d_val,size);
//--------------sleep for a while --------------------------
*h_flag=false;
cudaMemcpy(d_flag,h_flag,1,cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
gpuErrchk( cudaPeekAtLastError() );
printf("END\n");
}
where the kernel stopMe keeps running until someone from the host side sets up the flag to false. Note that your kernel could be much more complicated than this and the effort to synchronize all threads in order to execute the return could be much more than this (and can affect performance). Hope this helped.
More info here