1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 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 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
|
//------------------------------------------------------------------------------
// AxB_dot3_phase3_spdn.cu
//------------------------------------------------------------------------------
// This CUDA kernel produces the semi-ring product of two
// sparse matrices of types T_A and T_B and common index space size n, to a
// output matrix of type T_C. The matrices are sparse, with different numbers
// of non-zeros and different sparsity patterns.
// ie. we want to produce C = A'*B in the sense of the given semi-ring.
// This version uses an entire threadblock to compute each C(i,j) dot product.
// Both the grid and block are 1D, so blockDim.x is the # threads in a
// threadblock, and the # of threadblocks is grid.x
// int64_t start <- start of vector pairs for this kernel
// int64_t end <- end of vector pairs for this kernel
// int64_t *Bucket <- array of pair indices for all kernels
// matrix<T_C> *C <- result matrix
// matrix<T_M> *M <- mask matrix
// matrix<T_A> *A <- input matrix A
// matrix<T_B> *B <- input matrix B
#pragma once
#include <limits>
#include <cstdint>
#include <cooperative_groups.h>
#include "GB_cuda_kernel.h"
#include "GB_hash.h"
#include "GB_hyper_hash_lookup.h"
// Using tile size fixed at compile time, we don't need shared memory
#define tile_sz 32
using namespace cooperative_groups;
// FIXME: for the ANY monoid, GB_reduce_sum becomes trivial.
// or, if terminal condition is hit.
// FIXME: move this out of here, to share it
template< typename T, int warp_sz>
__device__ __inline__
T GB_reduce_sum(thread_block_tile<warp_sz> g, T val)
{
// Each iteration halves the number of active threads
// Each thread adds its partial sum[i] to sum[lane+i]
// Temporary T is necessary to handle arbirary ops
#pragma unroll
for (int i = warp_sz >> 1; i > 0; i >>= 1)
{
T next = g.shfl_down( val, i);
GB_ADD( val, val, next );
}
return val;
}
template< typename T, int warp_sz>
__device__ __inline__
T reduce_plus(thread_block_tile<warp_sz> g, T val)
{
// Each iteration halves the number of active threads
// Each thread adds its partial sum[i] to sum[lane+i]
#pragma unroll
for (int i = warp_sz >> 1; i > 0; i >>= 1)
{
val += g.shfl_down( val, i) ;
}
return val; // note: only thread 0 will return full sum and flag value
}
template<
typename T_C, typename T_A, typename T_B,
typename T_Z, typename T_X, typename T_Y,
uint64_t srcode>
__global__ void AxB_dot3_phase3_spdn
(
int64_t start,
int64_t end,
int64_t *Bucket, // do the work in Bucket [start:end-1]
GrB_Matrix C,
GrB_Matrix M,
GrB_Matrix A,
GrB_Matrix B,
int sz // FIXME: unused
)
{
// TODO: Figure out how to use graphblas-specific INFINITY macro
#ifndef INFINITY
#define INFINITY std::numeric_limits<T_C>::max()
#endif
const T_A *__restrict__ Ax = (T_A *)A->x ;
const T_B *__restrict__ Bx = (T_B *)B->x ;
T_C *__restrict__ Cx = (T_C *)C->x ;
int64_t *__restrict__ Ci = C->i ;
const int64_t *__restrict__ Mi = M->i ;
#if GB_M_IS_HYPER
const int64_t *__restrict__ Mh = M->h ;
#endif
#if GB_A_IS_HYPER || GB_A_IS_SPARSE
const int64_t *__restrict__ Ai = A->i ;
const int64_t *__restrict__ Ap = A->p ;
#endif
#if GB_A_IS_BITMAP
const int8_t *__restrict__ Ab = A->b ;
#endif
#if GB_B_IS_HYPER || GB_B_IS_SPARSE
const int64_t *__restrict__ Bi = B->i ;
const int64_t *__restrict__ Bp = B->p ;
#endif
#if GB_B_IS_BITMAP
const int8_t *__restrict__ Bb = B->b ;
#endif
#if GB_A_IS_HYPER
const int64_t *__restrict__ A_Yp = A->Y->p ;
const int64_t *__restrict__ A_Yi = A->Y->i ;
const int64_t *__restrict__ A_Yx = (int64_t *) A->Y->x ;
const int64_t A_hash_bits = A->Y->vdim - 1 ;
#endif
#if GB_B_IS_HYPER
const int64_t *__restrict__ B_Yp = B->Y->p ;
const int64_t *__restrict__ B_Yi = B->Y->i ;
const int64_t *__restrict__ B_Yx = (int64_t *) B->Y->x ;
const int64_t B_hash_bits = B->Y->vdim - 1 ;
#endif
// zombie count
int64_t zc = 0;
int64_t pair_id;
thread_block_tile<tile_sz> tile = tiled_partition<tile_sz>( this_thread_block());
int all_in_one = ( (end - start) == (M->p)[(M->nvec)] ) ;
// Main loop over pairs
int64_t kk ;
for (kk = start+ blockIdx.x; // warp per C(i,j)=A(:,i)'*B(:,j) dot product
kk < end;
kk += gridDim.x )
{
pair_id = all_in_one ? kk : Bucket [kk] ;
int64_t i = Mi[pair_id];
int64_t k = Ci[pair_id] >> 4;
// j = k or j = Mh [k] if C and M are hypersparse
#if GB_M_IS_HYPER
int64_t j = Mh [k] ;
#else
int64_t j = k ;
#endif
// find A(:,i)
int64_t pA, pA_end ;
#if GB_A_IS_HYPER
GB_hyper_hash_lookup (Ap, A_Yp, A_Yi, A_Yx, A_hash_bits,
i, &pA, &pA_end) ;
#elif GB_A_IS_SPARSE
pA = Ap[i] ;
pA_end = Ap[i+1] ;
#else
// A is bitmap or full
pA = A->vlen * i ;
pA_end = pA + i ;
#endif
GB_DECLAREA (aki) ;
GB_DECLAREB (bkj) ;
#if !GB_C_ISO
// T_Z cij = GB_IDENTITY ;
GB_DECLARE_MONOID_IDENTITY (cij) ;
#endif
int cij_exists = 0 ; // FIXME: make a bool
// find B(:,j)
int64_t pB, pB_end ;
#if GB_B_IS_HYPER
GB_hyper_hash_lookup (Bp, B_Yp, B_Yi, B_Yx, B_hash_bits,
j, &pB, &pB_end) ;
#elif GB_B_IS_SPARSE
pB = Bp[j] ;
pB_end = Bp[j+1] ;
#else
// B is bitmap or full
pB = B->vlen * j ;
pB_end = pB + j ;
#endif
//----------------------------------------------------------------------
// compute C(i,j) = A(:,i)'*B(:,j) using the entire threadblock
//----------------------------------------------------------------------
#if ( GB_A_IS_FULL )
{
// int64_t bjnz = pB_end - pB ; // bjnz = nnz (B (:,j))
// if (bjnz > 0) // will always be >= 128
{
//--------------------------------------------------------------
// A is full and B is sparse/hyper
//--------------------------------------------------------------
cij_exists = true ;
for (int64_t p = pB + threadIdx.x ; p < pB_end ; p += blockDim.x)
{
int64_t k = Bi [p] ; // next row index of B(:,j)
// cij += A(k,i) * B(k,j)
GB_GETA ( aki, Ax, pA+k ) ; // aki = A(k,i)
GB_GETB ( bkj, Bx, p ) ; // bkj = B(k,j)
GB_MULTADD ( cij, aki, bkj, i, k, j ) ; // cij += aki * bkj
GB_DOT_TERMINAL (cij) ; // break if cij == terminal
}
}
}
#elif ( GB_A_IS_BITMAP )
{
//------------------------------------------------------------------
// A is bitmap and B is sparse/hyper
//------------------------------------------------------------------
for (int64_t p = pB + threadIdx.x ; p < pB_end ; p += blockDim.x)
{
int64_t k = Bi [p] ; // next row index of B(:,j)
if (Ab [pA+k]) // check if A(k,i) exists
{
// cij += A(k,i) * B(k,j)
GB_DOT_MERGE (pA+k, p) ;
GB_DOT_TERMINAL (cij) ; // break if cij == terminal
}
}
}
#elif ( GB_B_IS_FULL )
{
// int64_t ainz = pA_end - pA ; // ainz = nnz (A (:,i))
// if (ainz > 0) // will always be >= 128
{
//--------------------------------------------------------------
// A is sparse/hyper and B is full
//--------------------------------------------------------------
cij_exists = true ;
for (int64_t p = pA + threadIdx.x ; p < pA_end ; p += blockDim.x)
{
int64_t k = Ai [p] ; // next row index of A(:,i)
// cij += A(k,i) * B(k,j)
GB_GETA ( aki, Ax, p ) ; // aki = A(i,k)
GB_GETB ( bkj, Bx, pB+k) ; // bkj = B(j,k)
GB_MULTADD ( cij, aki, bkj, i, k, j) ; // cij += aik * bjk
GB_DOT_TERMINAL (cij) ; // break if cij == terminal
}
}
}
#elif ( GB_B_IS_BITMAP )
{
//------------------------------------------------------------------
// A is sparse/hyper and B is bitmap
//------------------------------------------------------------------
for (int64_t p = pA + threadIdx.x ; p < pA_end ; p += blockDim.x)
{
int64_t k = Ai [p] ; // next row index of A(:,i)
if (Bb [pB+k]) // check if B(k,j) exists
{
// cij += A(k,i) * B(k,j)
GB_DOT_MERGE (p, pB+k) ;
GB_DOT_TERMINAL (cij) ; // break if cij == terminal
}
}
}
#endif
GB_CIJ_EXIST_POSTCHECK
//----------------------------------------------------------------------
// reduce sum per-thread values to a single scalar, get OR of flag
//----------------------------------------------------------------------
/*
if (threadIdx.x == 0)
{
printf ("reduce %d : %d exists = %d\n", b, cij, cij_exists) ;
}
__syncthreads();
*/
// Do vote here for control.
cij_exists = tile.any (cij_exists) ;
tile.sync ( ) ;
#if !GB_C_ISO
if (cij_exists)
{
cij = GB_reduce_sum<T_Z, tile_sz>( tile, cij );
}
#endif
// write result for this block to global mem
if (threadIdx.x == 0)
{
if (cij_exists)
{
GB_PUTC ( Cx[pair_id]=(T_C)cij ) ;
Ci[pair_id] = i ;
}
else
{
zc++;
Ci[pair_id]=GB_FLIP (i) ;
}
}
//__syncthreads();
}
//--------------------------------------------------------------------------
if(threadIdx.x ==0 && zc > 0)
{
// printf("warp %d zombie count = %d, nzombies = %d\n", blockIdx.x, zc, C->nzombies);
atomicAdd( (unsigned long long int*)&(C->nzombies), (unsigned long long int)zc);
// printf(" Czombie = %lld\n",C->nzombies);
}
//__syncthreads();
}
|