Actual source code: vecscattercuda.cu
petsc-3.7.5 2017-01-01
1: /*
2: Implements the various scatter operations on cuda vectors
3: */
5: #define PETSC_SKIP_SPINLOCK
7: #include <petscconf.h>
8: #include <petsc/private/vecimpl.h> /*I "petscvec.h" I*/
9: #include <../src/vec/vec/impls/dvecimpl.h>
10: #include <../src/vec/vec/impls/seq/seqcuda/cudavecimpl.h>
12: #include <cuda_runtime.h>
16: PetscErrorCode VecScatterCUDAIndicesCreate_StoS(PetscInt n,PetscInt toFirst,PetscInt fromFirst,PetscInt toStep, PetscInt fromStep,PetscInt *tslots, PetscInt *fslots,PetscCUDAIndices *ci) {
18: PetscCUDAIndices cci;
19: VecScatterCUDAIndices_StoS stos_scatter;
20: cudaError_t err;
21: cudaStream_t stream;
22: PetscInt *intVecGPU;
23: int device;
24: cudaDeviceProp props;
27: cci = new struct _p_PetscCUDAIndices;
28: stos_scatter = new struct _p_VecScatterCUDAIndices_StoS;
30: /* create the "from" indices */
31: stos_scatter->fslots = 0;
32: stos_scatter->fromFirst = 0;
33: stos_scatter->fromStep = 0;
34: if (n) {
35: if (fslots) {
36: /* allocate GPU memory for the to-slots */
37: err = cudaMalloc((void **)&intVecGPU,n*sizeof(PetscInt));CHKERRCUDA(err);
38: err = cudaMemcpy(intVecGPU,fslots,n*sizeof(PetscInt),cudaMemcpyHostToDevice);CHKERRCUDA(err);
40: /* assign the pointer to the struct */
41: stos_scatter->fslots = intVecGPU;
42: stos_scatter->fromMode = VEC_SCATTER_CUDA_GENERAL;
43: } else if (fromStep) {
44: stos_scatter->fromFirst = fromFirst;
45: stos_scatter->fromStep = fromStep;
46: stos_scatter->fromMode = VEC_SCATTER_CUDA_STRIDED;
47: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must provide fslots or fromStep.");
48: }
50: /* create the "to" indices */
51: stos_scatter->tslots = 0;
52: stos_scatter->toFirst = 0;
53: stos_scatter->toStep = 0;
54: if (n) {
55: if (tslots) {
56: /* allocate GPU memory for the to-slots */
57: err = cudaMalloc((void **)&intVecGPU,n*sizeof(PetscInt));CHKERRCUDA(err);
58: err = cudaMemcpy(intVecGPU,tslots,n*sizeof(PetscInt),cudaMemcpyHostToDevice);CHKERRCUDA(err);
60: /* assign the pointer to the struct */
61: stos_scatter->tslots = intVecGPU;
62: stos_scatter->toMode = VEC_SCATTER_CUDA_GENERAL;
63: } else if (toStep) {
64: stos_scatter->toFirst = toFirst;
65: stos_scatter->toStep = toStep;
66: stos_scatter->toMode = VEC_SCATTER_CUDA_STRIDED;
67: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must provide tslots or toStep.");
68: }
70: /* allocate the stream variable */
71: err = cudaStreamCreate(&stream);CHKERRCUDA(err);
72: stos_scatter->stream = stream;
74: /* the number of indices */
75: stos_scatter->n = n;
77: /* get the maximum number of coresident thread blocks */
78: cudaGetDevice(&device);
79: cudaGetDeviceProperties(&props, device);
80: stos_scatter->MAX_CORESIDENT_THREADS = props.maxThreadsPerMultiProcessor;
81: if (props.major>=3) {
82: stos_scatter->MAX_BLOCKS = 16*props.multiProcessorCount;
83: } else {
84: stos_scatter->MAX_BLOCKS = 8*props.multiProcessorCount;
85: }
87: /* assign the indices */
88: cci->scatter = (VecScatterCUDAIndices_StoS)stos_scatter;
89: cci->scatterType = VEC_SCATTER_CUDA_STOS;
90: *ci = cci;
91: return(0);
92: }
96: PetscErrorCode VecScatterCUDAIndicesCreate_PtoP(PetscInt ns,PetscInt *sendIndices,PetscInt nr,PetscInt *recvIndices,PetscCUDAIndices *ci)
97: {
98: PetscCUDAIndices cci;
99: VecScatterCUDAIndices_PtoP ptop_scatter;
102: cci = new struct _p_PetscCUDAIndices;
103: ptop_scatter = new struct _p_VecScatterCUDAIndices_PtoP;
105: /* this calculation assumes that the input indices are sorted */
106: ptop_scatter->ns = sendIndices[ns-1]-sendIndices[0]+1;
107: ptop_scatter->sendLowestIndex = sendIndices[0];
108: ptop_scatter->nr = recvIndices[nr-1]-recvIndices[0]+1;
109: ptop_scatter->recvLowestIndex = recvIndices[0];
111: /* assign indices */
112: cci->scatter = (VecScatterCUDAIndices_PtoP)ptop_scatter;
113: cci->scatterType = VEC_SCATTER_CUDA_PTOP;
115: *ci = cci;
116: return(0);
117: }
121: PetscErrorCode VecScatterCUDAIndicesDestroy(PetscCUDAIndices *ci)
122: {
124: if (*ci) {
125: if ((*ci)->scatterType == VEC_SCATTER_CUDA_PTOP) {
126: delete (VecScatterCUDAIndices_PtoP)(*ci)->scatter;
127: (*ci)->scatter = 0;
128: } else {
129: cudaError_t err;
130: VecScatterCUDAIndices_StoS stos_scatter = (VecScatterCUDAIndices_StoS)(*ci)->scatter;
131: if (stos_scatter->fslots) {
132: err = cudaFree(stos_scatter->fslots);CHKERRCUDA(err);
133: stos_scatter->fslots = 0;
134: }
136: /* free the GPU memory for the to-slots */
137: if (stos_scatter->tslots) {
138: err = cudaFree(stos_scatter->tslots);CHKERRCUDA(err);
139: stos_scatter->tslots = 0;
140: }
142: /* free the stream variable */
143: if (stos_scatter->stream) {
144: err = cudaStreamDestroy(stos_scatter->stream);CHKERRCUDA(err);
145: stos_scatter->stream = 0;
146: }
147: delete stos_scatter;
148: (*ci)->scatter = 0;
149: }
150: delete *ci;
151: *ci = 0;
152: }
153: return(0);
154: }
156: /* Insert operator */
157: class Insert {
158: public:
159: __device__ PetscScalar operator() (PetscScalar a,PetscScalar b) const {
160: return a;
161: }
162: };
164: /* Add operator */
165: class Add {
166: public:
167: __device__ PetscScalar operator() (PetscScalar a,PetscScalar b) const {
168: return a+b;
169: }
170: };
172: /* Add operator */
173: class Max {
174: public:
175: __device__ PetscScalar operator() (PetscScalar a,PetscScalar b) const {
176: return PetscMax(PetscRealPart(a),PetscRealPart(b));
177: }
178: };
180: /* Sequential general to sequential general GPU kernel */
181: template<class OPERATOR>
182: __global__ void VecScatterCUDA_SGtoSG_kernel(PetscInt n,PetscInt *xind,PetscScalar *x,PetscInt *yind,PetscScalar *y,OPERATOR OP) {
183: const int tidx = blockIdx.x*blockDim.x + threadIdx.x;
184: const int grid_size = gridDim.x * blockDim.x;
185: for (int i = tidx; i < n; i += grid_size) {
186: y[yind[i]] = OP(x[xind[i]],y[yind[i]]);
187: }
188: }
190: /* Sequential general to sequential strided GPU kernel */
191: template<class OPERATOR>
192: __global__ void VecScatterCUDA_SGtoSS_kernel(PetscInt n,PetscInt *xind,PetscScalar *x,PetscInt toFirst,PetscInt toStep,PetscScalar *y,OPERATOR OP) {
193: const int tidx = blockIdx.x*blockDim.x + threadIdx.x;
194: const int grid_size = gridDim.x * blockDim.x;
195: for (int i = tidx; i < n; i += grid_size) {
196: y[toFirst+i*toStep] = OP(x[xind[i]],y[toFirst+i*toStep]);
197: }
198: }
200: /* Sequential strided to sequential strided GPU kernel */
201: template<class OPERATOR>
202: __global__ void VecScatterCUDA_SStoSS_kernel(PetscInt n,PetscInt fromFirst,PetscInt fromStep,PetscScalar *x,PetscInt toFirst,PetscInt toStep,PetscScalar *y,OPERATOR OP) {
203: const int tidx = blockIdx.x*blockDim.x + threadIdx.x;
204: const int grid_size = gridDim.x * blockDim.x;
205: for (int i = tidx; i < n; i += grid_size) {
206: y[toFirst+i*toStep] = OP(x[fromFirst+i*fromStep],y[toFirst+i*toStep]);
207: }
208: }
210: /* Sequential strided to sequential general GPU kernel */
211: template<class OPERATOR>
212: __global__ void VecScatterCUDA_SStoSG_kernel(PetscInt n,PetscInt fromFirst,PetscInt fromStep,PetscScalar *x,PetscInt *yind,PetscScalar *y,OPERATOR OP) {
213: const int tidx = blockIdx.x*blockDim.x + threadIdx.x;
214: const int grid_size = gridDim.x * blockDim.x;
215: for (int i = tidx; i < n; i += grid_size) {
216: y[yind[i]] = OP(x[fromFirst+i*fromStep],y[yind[i]]);
217: }
218: }
220: template<class OPERATOR>
221: void VecScatterCUDA_StoS_Dispatcher(PetscScalar *xarray,PetscScalar *yarray,PetscCUDAIndices ci,ScatterMode mode,OPERATOR OP) {
223: PetscInt nBlocks=0,nThreads=128;
224: VecScatterCUDAIndices_StoS stos_scatter = (VecScatterCUDAIndices_StoS)ci->scatter;
226: nBlocks=(int)ceil(((float) stos_scatter->n)/((float) nThreads))+1;
227: if (nBlocks>stos_scatter->MAX_CORESIDENT_THREADS/nThreads) {
228: nBlocks = stos_scatter->MAX_CORESIDENT_THREADS/nThreads;
229: }
230: dim3 block(nThreads,1,1);
231: dim3 grid(nBlocks,1,1);
233: if (mode == SCATTER_FORWARD) {
234: if (stos_scatter->fromMode == VEC_SCATTER_CUDA_GENERAL && stos_scatter->toMode == VEC_SCATTER_CUDA_GENERAL) {
235: VecScatterCUDA_SGtoSG_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->fslots,xarray,stos_scatter->tslots,yarray,OP);
236: } else if (stos_scatter->fromMode == VEC_SCATTER_CUDA_GENERAL && stos_scatter->toMode == VEC_SCATTER_CUDA_STRIDED) {
237: VecScatterCUDA_SGtoSS_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->fslots,xarray,stos_scatter->toFirst,stos_scatter->toStep,yarray,OP);
238: } else if (stos_scatter->fromMode == VEC_SCATTER_CUDA_STRIDED && stos_scatter->toMode == VEC_SCATTER_CUDA_STRIDED) {
239: VecScatterCUDA_SStoSS_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->fromFirst,stos_scatter->fromStep,xarray,stos_scatter->toFirst,stos_scatter->toStep,yarray,OP);
240: } else if (stos_scatter->fromMode == VEC_SCATTER_CUDA_STRIDED && stos_scatter->toMode == VEC_SCATTER_CUDA_GENERAL) {
241: VecScatterCUDA_SStoSG_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->fromFirst,stos_scatter->fromStep,xarray,stos_scatter->tslots,yarray,OP);
242: }
243: } else {
244: if (stos_scatter->toMode == VEC_SCATTER_CUDA_GENERAL && stos_scatter->fromMode == VEC_SCATTER_CUDA_GENERAL) {
245: VecScatterCUDA_SGtoSG_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->tslots,xarray,stos_scatter->fslots,yarray,OP);
246: } else if (stos_scatter->toMode == VEC_SCATTER_CUDA_GENERAL && stos_scatter->fromMode == VEC_SCATTER_CUDA_STRIDED) {
247: VecScatterCUDA_SGtoSS_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->tslots,xarray,stos_scatter->fromFirst,stos_scatter->fromStep,yarray,OP);
248: } else if (stos_scatter->toMode == VEC_SCATTER_CUDA_STRIDED && stos_scatter->fromMode == VEC_SCATTER_CUDA_STRIDED) {
249: VecScatterCUDA_SStoSS_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->toFirst,stos_scatter->toStep,xarray,stos_scatter->fromFirst,stos_scatter->fromStep,yarray,OP);
250: } else if (stos_scatter->toMode == VEC_SCATTER_CUDA_STRIDED && stos_scatter->fromMode == VEC_SCATTER_CUDA_GENERAL) {
251: VecScatterCUDA_SStoSG_kernel<<<grid,block,0,stos_scatter->stream>>>(stos_scatter->n,stos_scatter->toFirst,stos_scatter->toStep,xarray,stos_scatter->fslots,yarray,OP);
252: }
253: }
254: }
258: PetscErrorCode VecScatterCUDA_StoS(Vec x,Vec y,PetscCUDAIndices ci,InsertMode addv,ScatterMode mode)
259: {
260: PetscErrorCode ierr;
261: PetscScalar *xarray,*yarray;
262: VecScatterCUDAIndices_StoS stos_scatter = (VecScatterCUDAIndices_StoS)ci->scatter;
263: cudaError_t err;
266: VecCUDAAllocateCheck(x);
267: VecCUDAAllocateCheck(y);
268: VecCUDAGetArrayRead(x,&xarray);
269: VecCUDAGetArrayReadWrite(y,&yarray);
270: if (stos_scatter->n) {
271: if (addv == INSERT_VALUES)
272: VecScatterCUDA_StoS_Dispatcher(xarray,yarray,ci,mode,Insert());
273: else if (addv == ADD_VALUES)
274: VecScatterCUDA_StoS_Dispatcher(xarray,yarray,ci,mode,Add());
275: else if (addv == MAX_VALUES)
276: VecScatterCUDA_StoS_Dispatcher(xarray,yarray,ci,mode,Max());
277: else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Wrong insert option");
278: err = cudaGetLastError();CHKERRCUDA(err);
279: err = cudaStreamSynchronize(stos_scatter->stream);CHKERRCUDA(err);
280: }
281: VecCUDARestoreArrayRead(x,&xarray);
282: VecCUDARestoreArrayReadWrite(y,&yarray);
283: return(0);
284: }