ewGpuNode.cu 17.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
#include "ewGpuNode.cuh"
#include "ewCudaKernels.cuh"

CGpuNode::CGpuNode() {

	pitch = 0;
	copied = true;

	/* TODO: make dynamic */
	num_virtual_gpus = 4;
	num_real_gpus = 2;

	vgpus = new VGpu[num_virtual_gpus];
14
	gpus = new Gpu[num_real_gpus];
15 16 17

	cudaMallocHost( &extend, num_virtual_gpus * sizeof(int4) );

18 19 20 21 22 23 24 25 26 27 28 29
	for( int j = 0; j < num_real_gpus; j++ ) {

		cudaSetDevice( j );

		gpus[j].id = j;

		for( int i = 0; i < gpus[j].NEVENTS; i++ ) {
			cudaEventCreate( &(gpus[j].evtStart[i]) );
			cudaEventCreate( &(gpus[j].evtEnd[i]) );
			gpus[j].dur[i] = 0.0f;
		}

30 31 32 33 34 35 36 37 38
	}

	for( int j = 0; j < num_virtual_gpus; j++ ) {

		VGpu& vgpu = vgpus[j];

		vgpu.data.devID = j;
		vgpu.data.devNum = num_virtual_gpus;

39 40 41 42
		vgpu.dev = &(gpus[j % num_real_gpus]);

		vgpu.dev->maxId = j / num_real_gpus;
		vgpu.relId = j / num_real_gpus;
43

44
		cudaSetDevice( vgpu.dev->id );
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

		for( int i = 0; i < vgpu.NSTREAMS; i++ ) {
			cudaStreamCreate( &(vgpu.stream[i]) );
		}

		cudaEventCreate( &vgpu.evtSync );
	}

	for( int j = 0; j < num_real_gpus - 1; j++ ) {

		int peerAccess = 0;
		cudaDeviceCanAccessPeer( &peerAccess, j, j + 1 );

		printf_v("GPU #%u can access GPU #%u: %u\n", j, j + 1, peerAccess);
		if( peerAccess ) {
			cudaSetDevice( j );
			cudaDeviceEnablePeerAccess( j + 1, 0 );
		}

		cudaDeviceCanAccessPeer( &peerAccess, j + 1, j );

		printf_v("GPU #%u can access GPU #%u: %u\n", j + 1, j, peerAccess);
		if( peerAccess ) {
			cudaSetDevice( j + 1 );
			cudaDeviceEnablePeerAccess( j, 0 );
		}

	}

	memset( extend, 0, num_virtual_gpus * sizeof(int4) );
}

CGpuNode::~CGpuNode() {

	for( int j = 0; j < num_virtual_gpus; j++ ) {

		VGpu& vgpu = vgpus[j];

83
		cudaSetDevice( vgpu.dev->id );
84 85 86 87 88

		for( int i = 0; i < vgpu.NSTREAMS; i++ ) {
			cudaStreamDestroy( vgpu.stream[i] );
		}

89 90 91 92 93 94 95 96 97 98 99
		cudaEventDestroy( vgpu.evtSync );

	}

	for( int j = 0; j < num_real_gpus; j++ ) {

		cudaSetDevice( j );

		for( int i = 0; i < gpus[j].NEVENTS; i++ ) {
			cudaEventDestroy( gpus[j].evtStart[i] );
			cudaEventDestroy( gpus[j].evtEnd[i] );
100 101
		}

102
		cudaDeviceReset();
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
}

int CGpuNode::mallocMem() {

	CArrayNode::mallocMem();

	Params& dp = params;

	/* fill in some fields here */
	dp.nI = NLon;
	dp.nJ = NLat;
	dp.sshArrivalThreshold = Par.sshArrivalThreshold;
	dp.sshClipThreshold = Par.sshClipThreshold;
	dp.sshZeroThreshold = Par.sshZeroThreshold;
	dp.lpad = 0;

	size_t nJ_aligned = dp.nJ + dp.lpad;

	init_vgpus();

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

		int ghost = vgpu.gb + vgpu.gt;

132
		cudaSetDevice( vgpu.dev->id );
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

		/* arrays that need ghost zones must add 2 to vgpu.size */
		/* 2-dim */
		CUDA_CALL( cudaMallocPitch( &(data.d), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.fM), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.fN), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.cR1), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.cR2), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.cR4), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.tArr), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		/* TODO: cR3, cR5 for coriolis */

		CUDA_CALL( cudaMallocPitch( &(data.h), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );
		CUDA_CALL( cudaMallocPitch( &(data.hMax), &pitch, nJ_aligned * sizeof(float), vgpu.size + ghost ) );

		/* 1-dim */
		CUDA_CALL( cudaMalloc( &(data.cR6), dp.nJ * sizeof(float) ) );
		CUDA_CALL( cudaMalloc( &(data.cB1), (vgpu.size + ghost) * sizeof(float) ) );
		CUDA_CALL( cudaMalloc( &(data.cB2), dp.nJ * sizeof(float) ) );
		CUDA_CALL( cudaMalloc( &(data.cB3), (vgpu.size + ghost) * sizeof(float) ) );
		CUDA_CALL( cudaMalloc( &(data.cB4), dp.nJ * sizeof(float) ) );

		/* data field to store return values like calculated iMin and iMax values */
		CUDA_CALL( cudaMalloc( &(data.extend), sizeof(int4) ) );

		/* assign some parameters */
		data.params.nI = vgpu.getRel( vgpu.end );
		data.params.nJ = dp.nJ;
		data.params.sshArrivalThreshold = dp.sshArrivalThreshold;
		data.params.sshClipThreshold = dp.sshClipThreshold;
		data.params.sshZeroThreshold = dp.sshZeroThreshold;
		data.params.lpad = dp.lpad;

		/* TODO: make sure that pitch is a multiple of 4 and the same for each cudaMallocPitch() call */
		data.params.pI = pitch / sizeof(float);
	}

	return 0;
}

int CGpuNode::copyToGPU() {

	Params& dp = params;

	/* fill in further fields here */
	dp.iMin = Imin;
	dp.iMax = Imax;
	dp.jMin = Jmin;
	dp.jMax = Jmax;

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

		/* special treatment because of needed ghost zones */
		int off = (vgpu.off - vgpu.gt - 1);
		int ghost = vgpu.gb + vgpu.gt;

192
		cudaSetDevice( vgpu.dev->id );
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

		/* FIXME: should not work correctly */
		/* add offset to data.d to guarantee alignment: data.d + LPAD */
		/* 2-dim */
		CUDA_CALL( cudaMemcpy2D( data.d + dp.lpad, pitch, d + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.fM + dp.lpad, pitch, fM + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.fN + dp.lpad, pitch, fN + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.cR1 + dp.lpad, pitch, cR1 + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.cR2 + dp.lpad, pitch, cR2 + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.cR4 + dp.lpad, pitch, cR4 + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.tArr + dp.lpad, pitch, tArr + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.h + dp.lpad, pitch, h + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy2D( data.hMax + dp.lpad, pitch, hMax + off * dp.nJ, dp.nJ * sizeof(float), dp.nJ * sizeof(float), vgpu.size + ghost, cudaMemcpyHostToDevice ) );

		/* FIXME: move global variables into data structure */
		/* 1-dim */
		CUDA_CALL( cudaMemcpy( data.cR6, R6, dp.nJ * sizeof(float), cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy( data.cB1, C1 + off, (vgpu.size + ghost) * sizeof(float), cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy( data.cB2, C2, dp.nJ * sizeof(float), cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy( data.cB3, C3 + off, (vgpu.size + ghost) * sizeof(float), cudaMemcpyHostToDevice ) );
		CUDA_CALL( cudaMemcpy( data.cB4, C4, dp.nJ * sizeof(float), cudaMemcpyHostToDevice ) );

		data.params.jMin = dp.jMin;
		data.params.jMax = dp.jMax;
	}

	return 0;
}
int CGpuNode::copyFromGPU() {

	Params& dp = params;

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

		int off = (vgpu.off - 1) * dp.nJ;

232
		cudaSetDevice( vgpu.dev->id );
233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256

		CUDA_CALL( cudaMemcpy2D( hMax + off, dp.nJ * sizeof(float), data.hMax + (vgpu.gt)*data.params.pI + dp.lpad, pitch, dp.nJ * sizeof(float), vgpu.size, cudaMemcpyDeviceToHost ) );
		CUDA_CALL( cudaMemcpy2D( tArr + off, dp.nJ * sizeof(float), data.tArr + (vgpu.gt)*data.params.pI + dp.lpad, pitch, dp.nJ * sizeof(float), vgpu.size, cudaMemcpyDeviceToHost ) );

	}

	return 0;
}

int CGpuNode::copyIntermediate() {

	/* ignore copy requests if data already present on CPU side */
	if( copied )
		return 0;

	Params& dp = params;

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

		int off = (vgpu.off - 1) * dp.nJ;

257
		cudaSetDevice( vgpu.dev->id );
258 259 260 261 262 263 264 265 266 267 268

		CUDA_CALL( cudaMemcpy2D( h + off, dp.nJ * sizeof(float), data.h + (vgpu.gt) * data.params.pI + dp.lpad, pitch, dp.nJ * sizeof(float), vgpu.size, cudaMemcpyDeviceToHost ) );

	}

	/* copy finished */
	copied = true;

	return 0;
}

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
int CGpuNode::copyPOIs() {

	Params& dp = params;

	if( copied )
		return 0;

	VGpu *vgpu;

	for( int n = 0; n < NPOIs; n++ ) {

		int i = idxPOI[n] / dp.nJ + 1;
		int j = idxPOI[n] % dp.nJ + 1;

		for( int id = 0; id < num_virtual_gpus; id++ ) {

			if( vgpus[id].hasLine( i ) ) {
				vgpu = &(vgpus[id]);
				break;
			}

		}

		int id = vgpu->data.idx( vgpu->getRel(i), j );

294
		CUDA_CALL( cudaSetDevice( vgpu->dev->id ) )
295 296 297 298 299
		CUDA_CALL( cudaMemcpy( h + idxPOI[n], vgpu->data.h + dp.lpad + id, sizeof(float), cudaMemcpyDeviceToHost ) );
	}

	return 0;
}
300 301 302 303 304 305 306 307

int CGpuNode::freeMem() {

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

308
		cudaSetDevice( vgpu.dev->id );
309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327

		/* 2-dim */
		CUDA_CALL( cudaFree( data.d ) );
		CUDA_CALL( cudaFree( data.h ) );
		CUDA_CALL( cudaFree( data.hMax ) );
		CUDA_CALL( cudaFree( data.fM ) );
		CUDA_CALL( cudaFree( data.fN ) );
		CUDA_CALL( cudaFree( data.cR1 ) );
		CUDA_CALL( cudaFree( data.cR2 ) );
		CUDA_CALL( cudaFree( data.cR4 ) );
		CUDA_CALL( cudaFree( data.tArr ) );

		/* 1-dim */
		CUDA_CALL( cudaFree( data.cR6 ) );
		CUDA_CALL( cudaFree( data.cB1 ) );
		CUDA_CALL( cudaFree( data.cB2 ) );
		CUDA_CALL( cudaFree( data.cB3 ) );
		CUDA_CALL( cudaFree( data.cB4 ) );

328 329
		CUDA_CALL( cudaFree( data.extend ) );
	}
330

331
	cudaFreeHost( extend );
332

333
	for( int i = 0; i < num_real_gpus; i++ ) {
334

335 336
		float dur = 0.0f;
		for( int j = 0; j < 7; j++ ) {
337

338 339 340 341
			printf_v("GPU #%u, duration %u: %.3f\n", i, j, gpus[i].dur[j]);
			dur += gpus[i].dur[j];
		}
		printf_v("GPU #%u, duration total: %.3f\n", i, dur);
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
	}

	CArrayNode::freeMem();

	return 0;
}

int CGpuNode::run() {

	Params& dp = params;

	int nThreads = 256;
	int xThreads = 32;
	int yThreads = nThreads / xThreads;

	int4 glb_MinMax = {0,0,0,0};

	dp.mTime = Par.time;

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

		updateParams( vgpu );

		vgpu.nBlocks = ceil( (float)max(data.params.nI,dp.nJ) / (float)nThreads );
		vgpu.threads = dim3( xThreads, yThreads );
		vgpu.blocks = dim3( ceil( (float)dp.nJ / (float)xThreads ), ceil( (float)data.params.nI / (float)yThreads ) );

372 373 374 375 376 377
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );

		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[0] ) );

		if( isActive( vgpu ) ) {
378

379 380
			runWaveUpdateKernel<<<vgpu.blocks,vgpu.threads,0,vgpu.stream[0]>>>( data );
		}
381

382 383
		if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
				CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[0] ) );
384 385 386 387 388 389 390
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

391
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );
392

393 394
		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[1] ) );
395

396 397 398 399 400 401 402
		if( isActive( vgpu ) ) {

			runWaveBoundaryKernel<<<vgpu.nBlocks,nThreads,0,vgpu.stream[0]>>>( data );
		}

	if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
		CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[1] ) );
403 404 405 406 407 408
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];

409
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );
410

411 412
		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[5] ) );
413

414 415 416 417 418 419 420 421 422 423 424
		if( isActive( vgpu ) ) {

			if( i < num_virtual_gpus - 1 ) {
				int off = ( vgpu.getRel(vgpu.end) - 1 ) * vgpu.data.params.pI;
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i+1].data.h, vgpus[i+1].dev->id, vgpu.data.h + off, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0]) );
			}

			if( i > 0 ) {
				int off = ( vgpus[i-1].getRel(vgpus[i-1].end) ) * vgpus[i-1].data.params.pI;
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i-1].data.h + off, vgpus[i-1].dev->id, vgpu.data.h + vgpu.data.params.pI, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0] ) );
			}
425 426

		}
427 428 429

		if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[5] ) );
430 431 432 433 434 435 436 437 438 439 440

		cudaEventRecord( vgpu.evtSync, vgpu.stream[0] );
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];

		if( ! isActive(vgpu) )
			continue;

441
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );
442 443 444 445 446 447 448 449 450 451 452 453 454

		if( i < num_virtual_gpus - 1 )
			cudaStreamWaitEvent( vgpu.stream[0], vgpus[i+1].evtSync, 0 );

		if( i > 0 )
			cudaStreamWaitEvent( vgpu.stream[0], vgpus[i-1].evtSync, 0 );
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

455 456 457 458
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );

		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[2] ) );
459

460
		if( isActive( vgpu ) ) {
461

462 463 464 465 466
			runFluxUpdateKernel<<<vgpu.blocks,vgpu.threads,0,vgpu.stream[0]>>>( data );
		}

		if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
				CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[2] ) );
467 468 469 470 471 472 473
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

474 475 476 477
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );

		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[3] ) );
478

479
		if( isActive( vgpu ) ) {
480

481 482 483 484 485
			runFluxBoundaryKernel<<<vgpu.nBlocks,nThreads,0,vgpu.stream[0]>>>( data );
		}

	if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[3] ) );
486 487 488 489 490 491
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];

492
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );
493

494 495
		if( Par.verbose && vgpu.relId == 0 )
				CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[6] ) );
496

497 498 499 500 501 502 503 504 505 506 507 508 509
		if( isActive( vgpu ) ) {

			if( i < num_virtual_gpus - 1 ) {
				int off = ( vgpu.getRel(vgpu.end) - 1 ) * vgpu.data.params.pI;
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i+1].data.fN, vgpus[i+1].dev->id, vgpu.data.fN + off, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0]) );
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i+1].data.fM, vgpus[i+1].dev->id, vgpu.data.fM + off, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0]) );
			}

			if( i > 0 ) {
				int off = ( vgpus[i-1].getRel(vgpus[i-1].end) ) * vgpus[i-1].data.params.pI;
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i-1].data.fN + off, vgpus[i-1].dev->id, vgpu.data.fN + vgpu.data.params.pI, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0] ) );
				CUDA_CALL( cudaMemcpyPeerAsync( vgpus[i-1].data.fM + off, vgpus[i-1].dev->id, vgpu.data.fM + vgpu.data.params.pI, vgpu.dev->id, vgpu.data.params.pI * sizeof(float), vgpu.stream[0] ) );
			}
510 511 512

		}

513 514
		if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
				CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[6] ) );
515 516 517 518 519 520 521
	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

522 523 524 525
		CUDA_CALL( cudaSetDevice( vgpu.dev->id ) );

		if( Par.verbose && vgpu.relId == 0 )
			CUDA_CALL( cudaEventRecord( vgpu.dev->evtStart[4] ) );
526

527 528 529 530 531 532 533 534
		if( isActive( vgpu ) ) {

			runGridExtendKernel<<<vgpu.nBlocks,nThreads,0,vgpu.stream[0]>>>( data );
			CUDA_CALL( cudaMemcpyAsync( &(extend[i]), data.extend, sizeof(int4), cudaMemcpyDeviceToHost, vgpu.stream[0]) );
		}

		if( Par.verbose && vgpu.relId == vgpu.dev->maxId )
				CUDA_CALL( cudaEventRecord( vgpu.dev->evtEnd[4] ) );
535 536 537 538 539 540 541 542

	}

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		VGpu& vgpu = vgpus[i];
		KernelData& data = vgpu.data;

543
		cudaSetDevice( vgpu.dev->id );
544 545 546 547 548 549 550 551 552 553 554
		cudaDeviceSynchronize();
		CUDA_CALL( cudaMemset( data.extend, 0, sizeof(int4) ) );

		if( vgpu.hasLine( dp.iMin + 2 ) )
			glb_MinMax.x += extend[i].x;

		if( vgpu.hasLine( dp.iMax - 2 ) )
			glb_MinMax.y += extend[i].y;

		glb_MinMax.z += extend[i].z;
		glb_MinMax.w += extend[i].w;
555
	}
556

557 558 559 560 561 562 563 564 565 566
	if( Par.verbose ) {
		for( int i = 0; i < num_real_gpus; i++ ) {

			cudaSetDevice( i );

			float dur;
			for( int j = 0; j < 7; j++ ) {
				if( cudaEventElapsedTime( &dur, gpus[i].evtStart[j],  gpus[i].evtEnd[j]) == cudaSuccess )
					gpus[i].dur[j] += dur;
			}
567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
		}
	}

	memset( extend, 0, num_virtual_gpus * sizeof(int4) );

	if( glb_MinMax.x > 0 )
		Imin = dp.iMin = max( dp.iMin-1, 2 );

	if( glb_MinMax.y > 0 )
		Imax = dp.iMax = min( dp.iMax+1, dp.nI-1 );

	if( glb_MinMax.z > 0 )
		Jmin = dp.jMin = max( dp.jMin-1, 2 );

	if( glb_MinMax.w > 0 )
		Jmax = dp.jMax = min( dp.jMax+1, dp.nJ-1 );

	/* data has changed now -> copy becomes necessary */
	copied = false;

	return 0;
}

int CGpuNode::init_vgpus() {

	Params& dp = params;

	int nI_partial = dp.nI / num_virtual_gpus;
	int nI_mod = dp.nI % num_virtual_gpus;
	int off = 1;

	for( int i = 0; i < num_virtual_gpus; i++ ) {

		vgpus[i].size = nI_partial + (int)( i < nI_mod );
		vgpus[i].off = off;
		vgpus[i].end = vgpus[i].off + vgpus[i].size - 1;
		off += vgpus[i].size;

		vgpus[i].gt = 1 - (int)( i == 0 );
		vgpus[i].gb = 1 - (int)( i == num_virtual_gpus - 1 );

		printf_v("VGPU #%u: off=%u end=%u size=%u gt=%u gb=%u\n", i, vgpus[i].off, vgpus[i].end, vgpus[i].size, vgpus[i].gt, vgpus[i].gb);
	}

	return 0;
}

int CGpuNode::updateParams( VGpu& vgpu ) {

	Params& dp = vgpu.data.params;

	dp.mTime = params.mTime;
	dp.jMin = params.jMin;
	dp.jMax = params.jMax;

	if( params.iMin <= vgpu.end && params.iMax >= vgpu.off ) {

		/* range of virtual GPU must be handled */
		/* calculate relative values for iMin and iMax depending on current range */
		dp.iMin = max( 2, vgpu.getRel(params.iMin) );
		dp.iMax = min( vgpu.getRel( vgpu.end ), vgpu.getRel(params.iMax) );

		return 0;
	}

	return 1;
}

bool CGpuNode::isActive( VGpu& vgpu ) {

	return ( params.iMin <= vgpu.end && params.iMax >= vgpu.off );
}