Actual source code: bvorthogcuda.cu
slepc-3.15.2 2021-09-20
1: /*
2: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3: SLEPc - Scalable Library for Eigenvalue Problem Computations
4: Copyright (c) 2002-2021, Universitat Politecnica de Valencia, Spain
6: This file is part of SLEPc.
7: SLEPc is distributed under a 2-clause BSD license (see LICENSE).
8: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
9: */
10: /*
11: BV orthogonalization routines (CUDA)
12: */
14: #include <slepc/private/bvimpl.h>
15: #include <slepcblaslapack.h>
16: #include <slepccublas.h>
18: /*
19: BV_CleanCoefficients_CUDA - Sets to zero all entries of column j of the bv buffer
20: */
21: PetscErrorCode BV_CleanCoefficients_CUDA(BV bv,PetscInt j,PetscScalar *h)
22: {
24: PetscScalar *d_hh,*d_a;
25: PetscInt i;
26: cudaError_t cerr;
29: if (!h) {
30: VecCUDAGetArray(bv->buffer,&d_a);
31: d_hh = d_a + j*(bv->nc+bv->m);
32: cerr = cudaMemset(d_hh,0,(bv->nc+j)*sizeof(PetscScalar));CHKERRCUDA(cerr);
33: cerr = WaitForCUDA();CHKERRCUDA(cerr);
34: VecCUDARestoreArray(bv->buffer,&d_a);
35: } else { /* cpu memory */
36: for (i=0;i<bv->nc+j;i++) h[i] = 0.0;
37: }
38: return(0);
39: }
41: /*
42: BV_AddCoefficients_CUDA - Add the contents of the scratch (0-th column) of the bv buffer
43: into column j of the bv buffer
44: */
45: PetscErrorCode BV_AddCoefficients_CUDA(BV bv,PetscInt j,PetscScalar *h,PetscScalar *c)
46: {
48: PetscScalar *d_h,*d_c,sone=1.0;
49: PetscInt i;
50: PetscBLASInt idx=0,one=1;
51: cublasStatus_t cberr;
52: cublasHandle_t cublasv2handle;
55: if (!h) {
56: PetscCUBLASGetHandle(&cublasv2handle);
57: VecCUDAGetArray(bv->buffer,&d_c);
58: d_h = d_c + j*(bv->nc+bv->m);
59: PetscBLASIntCast(bv->nc+j,&idx);
60: PetscLogGpuTimeBegin();
61: cberr = cublasXaxpy(cublasv2handle,idx,&sone,d_c,one,d_h,one);CHKERRCUBLAS(cberr);
62: PetscLogGpuTimeEnd();
63: PetscLogGpuFlops(1.0*(bv->nc+j));
64: VecCUDARestoreArray(bv->buffer,&d_c);
65: } else { /* cpu memory */
66: for (i=0;i<bv->nc+j;i++) h[i] += c[i];
67: PetscLogFlops(1.0*(bv->nc+j));
68: }
69: return(0);
70: }
72: /*
73: BV_SetValue_CUDA - Sets value in row j (counted after the constraints) of column k
74: of the coefficients array
75: */
76: PetscErrorCode BV_SetValue_CUDA(BV bv,PetscInt j,PetscInt k,PetscScalar *h,PetscScalar value)
77: {
79: PetscScalar *d_h,*a;
80: cudaError_t cerr;
83: if (!h) {
84: VecCUDAGetArray(bv->buffer,&a);
85: d_h = a + k*(bv->nc+bv->m) + bv->nc+j;
86: cerr = cudaMemcpy(d_h,&value,sizeof(PetscScalar),cudaMemcpyHostToDevice);CHKERRCUDA(cerr);
87: PetscLogCpuToGpu(sizeof(PetscScalar));
88: cerr = WaitForCUDA();CHKERRCUDA(cerr);
89: VecCUDARestoreArray(bv->buffer,&a);
90: } else { /* cpu memory */
91: h[bv->nc+j] = value;
92: }
93: return(0);
94: }
96: /*
97: BV_SquareSum_CUDA - Returns the value h'*h, where h represents the contents of the
98: coefficients array (up to position j)
99: */
100: PetscErrorCode BV_SquareSum_CUDA(BV bv,PetscInt j,PetscScalar *h,PetscReal *sum)
101: {
102: PetscErrorCode ierr;
103: const PetscScalar *d_h;
104: PetscScalar dot;
105: PetscInt i;
106: PetscBLASInt idx=0,one=1;
107: cublasStatus_t cberr;
108: cublasHandle_t cublasv2handle;
111: if (!h) {
112: PetscCUBLASGetHandle(&cublasv2handle);
113: VecCUDAGetArrayRead(bv->buffer,&d_h);
114: PetscBLASIntCast(bv->nc+j,&idx);
115: PetscLogGpuTimeBegin();
116: cberr = cublasXdotc(cublasv2handle,idx,d_h,one,d_h,one,&dot);CHKERRCUBLAS(cberr);
117: PetscLogGpuTimeEnd();
118: PetscLogGpuFlops(2.0*(bv->nc+j));
119: *sum = PetscRealPart(dot);
120: VecCUDARestoreArrayRead(bv->buffer,&d_h);
121: } else { /* cpu memory */
122: *sum = 0.0;
123: for (i=0;i<bv->nc+j;i++) *sum += PetscRealPart(h[i]*PetscConj(h[i]));
124: PetscLogFlops(2.0*(bv->nc+j));
125: }
126: return(0);
127: }
129: #define X_AXIS 0
130: #define BLOCK_SIZE_X 64
131: #define TILE_SIZE_X 16 /* work to be done by any thread on axis x */
133: /*
134: Set the kernels grid dimensions
135: xcount: number of kernel calls needed for the requested size
136: */
137: PetscErrorCode SetGrid1D(PetscInt n, dim3 *dimGrid, dim3 *dimBlock,PetscInt *xcount)
138: {
139: PetscInt one=1;
140: PetscBLASInt card;
141: struct cudaDeviceProp devprop;
142: cudaError_t cerr;
145: *xcount = 1;
146: if (n>BLOCK_SIZE_X) {
147: dimBlock->x = BLOCK_SIZE_X;
148: dimGrid->x = (n+BLOCK_SIZE_X*TILE_SIZE_X-one)/BLOCK_SIZE_X*TILE_SIZE_X;
149: } else {
150: dimBlock->x = (n+TILE_SIZE_X-one)/TILE_SIZE_X;
151: dimGrid->x = one;
152: }
153: cerr = cudaGetDevice(&card);CHKERRCUDA(cerr);
154: cerr = cudaGetDeviceProperties(&devprop,card);CHKERRCUDA(cerr);
155: if (dimGrid->x>(unsigned)devprop.maxGridSize[X_AXIS]) {
156: *xcount = (dimGrid->x+devprop.maxGridSize[X_AXIS]-one)/devprop.maxGridSize[X_AXIS];
157: dimGrid->x = devprop.maxGridSize[X_AXIS];
158: }
159: return(0);
160: }
162: /* pointwise multiplication */
163: __global__ void PointwiseMult_kernel(PetscInt xcount,PetscScalar *a,const PetscScalar *b,PetscInt n)
164: {
165: PetscInt i,x;
167: x = xcount*gridDim.x*blockDim.x+blockIdx.x*blockDim.x*TILE_SIZE_X+threadIdx.x*TILE_SIZE_X;
168: for (i=x;i<x+TILE_SIZE_X&&i<n;i++) {
169: a[i] *= PetscRealPart(b[i]);
170: }
171: }
173: /* pointwise division */
174: __global__ void PointwiseDiv_kernel(PetscInt xcount,PetscScalar *a,const PetscScalar *b,PetscInt n)
175: {
176: PetscInt i,x;
178: x = xcount*gridDim.x*blockDim.x+blockIdx.x*blockDim.x*TILE_SIZE_X+threadIdx.x*TILE_SIZE_X;
179: for (i=x;i<x+TILE_SIZE_X&&i<n;i++) {
180: a[i] /= PetscRealPart(b[i]);
181: }
182: }
184: /*
185: BV_ApplySignature_CUDA - Computes the pointwise product h*omega, where h represents
186: the contents of the coefficients array (up to position j) and omega is the signature;
187: if inverse=TRUE then the operation is h/omega
188: */
189: PetscErrorCode BV_ApplySignature_CUDA(BV bv,PetscInt j,PetscScalar *h,PetscBool inverse)
190: {
191: PetscErrorCode ierr;
192: PetscScalar *d_h;
193: const PetscScalar *d_omega,*omega;
194: PetscInt i,xcount;
195: dim3 blocks3d, threads3d;
196: cudaError_t cerr;
199: if (!(bv->nc+j)) return(0);
200: if (!h) {
201: VecCUDAGetArray(bv->buffer,&d_h);
202: VecCUDAGetArrayRead(bv->omega,&d_omega);
203: SetGrid1D(bv->nc+j,&blocks3d,&threads3d,&xcount);
204: PetscLogGpuTimeBegin();
205: if (inverse) {
206: for (i=0;i<xcount;i++) {
207: PointwiseDiv_kernel<<<blocks3d,threads3d>>>(i,d_h,d_omega,bv->nc+j);
208: }
209: } else {
210: for (i=0;i<xcount;i++) {
211: PointwiseMult_kernel<<<blocks3d,threads3d>>>(i,d_h,d_omega,bv->nc+j);
212: }
213: }
214: cerr = cudaGetLastError();CHKERRCUDA(cerr);
215: PetscLogGpuTimeEnd();
216: PetscLogGpuFlops(1.0*(bv->nc+j));
217: cerr = WaitForCUDA();CHKERRCUDA(cerr);
218: VecCUDARestoreArrayRead(bv->omega,&d_omega);
219: VecCUDARestoreArray(bv->buffer,&d_h);
220: } else {
221: VecGetArrayRead(bv->omega,&omega);
222: if (inverse) for (i=0;i<bv->nc+j;i++) h[i] /= PetscRealPart(omega[i]);
223: else for (i=0;i<bv->nc+j;i++) h[i] *= PetscRealPart(omega[i]);
224: VecRestoreArrayRead(bv->omega,&omega);
225: PetscLogFlops(1.0*(bv->nc+j));
226: }
227: return(0);
228: }
230: /*
231: BV_SquareRoot_CUDA - Returns the square root of position j (counted after the constraints)
232: of the coefficients array
233: */
234: PetscErrorCode BV_SquareRoot_CUDA(BV bv,PetscInt j,PetscScalar *h,PetscReal *beta)
235: {
236: PetscErrorCode ierr;
237: const PetscScalar *d_h;
238: PetscScalar hh;
239: cudaError_t cerr;
242: if (!h) {
243: VecCUDAGetArrayRead(bv->buffer,&d_h);
244: cerr = cudaMemcpy(&hh,d_h+bv->nc+j,sizeof(PetscScalar),cudaMemcpyDeviceToHost);CHKERRCUDA(cerr);
245: PetscLogGpuToCpu(sizeof(PetscScalar));
246: cerr = WaitForCUDA();CHKERRCUDA(cerr);
247: BV_SafeSqrt(bv,hh,beta);
248: VecCUDARestoreArrayRead(bv->buffer,&d_h);
249: } else {
250: BV_SafeSqrt(bv,h[bv->nc+j],beta);
251: }
252: return(0);
253: }
255: /*
256: BV_StoreCoefficients_CUDA - Copy the contents of the coefficients array to an array dest
257: provided by the caller (only values from l to j are copied)
258: */
259: PetscErrorCode BV_StoreCoefficients_CUDA(BV bv,PetscInt j,PetscScalar *h,PetscScalar *dest)
260: {
261: PetscErrorCode ierr;
262: const PetscScalar *d_h,*d_a;
263: PetscInt i;
264: cudaError_t cerr;
267: if (!h) {
268: VecCUDAGetArrayRead(bv->buffer,&d_a);
269: d_h = d_a + j*(bv->nc+bv->m)+bv->nc;
270: cerr = cudaMemcpy(dest-bv->l,d_h,(j-bv->l)*sizeof(PetscScalar),cudaMemcpyDeviceToHost);CHKERRCUDA(cerr);
271: PetscLogGpuToCpu((j-bv->l)*sizeof(PetscScalar));
272: cerr = WaitForCUDA();CHKERRCUDA(cerr);
273: VecCUDARestoreArrayRead(bv->buffer,&d_a);
274: } else {
275: for (i=bv->l;i<j;i++) dest[i-bv->l] = h[bv->nc+i];
276: }
277: return(0);
278: }