time_cuda.cu (6466B)
1 /* 2 Copyright 2010-2011, D. E. Shaw Research. 3 All rights reserved. 4 5 Redistribution and use in source and binary forms, with or without 6 modification, are permitted provided that the following conditions are 7 met: 8 9 * Redistributions of source code must retain the above copyright 10 notice, this list of conditions, and the following disclaimer. 11 12 * Redistributions in binary form must reproduce the above copyright 13 notice, this list of conditions, and the following disclaimer in the 14 documentation and/or other materials provided with the distribution. 15 16 * Neither the name of D. E. Shaw Research nor the names of its 17 contributors may be used to endorse or promote products derived from 18 this software without specific prior written permission. 19 20 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 23 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 24 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 25 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 26 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 27 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 28 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 29 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 30 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 31 */ 32 /* 33 * CUDA test and timing harness for Random123 RNGs. Uses macros 34 * and util_expandtpl.h to "templatize" over all the different 35 * permutations of RNGs and NxW and R. 36 */ 37 38 #include "util_cuda.h" 39 40 #include "Random123/philox.h" 41 #include "Random123/threefry.h" 42 43 #include "time_misc.h" 44 #include "util_print.h" 45 46 #define KERNEL __global__ 47 #define get_global_id(i) (blockDim.x * blockIdx.x + threadIdx.x) 48 #include "time_random123.h" 49 50 #define TEST_TPL(NAME, N, W, R) \ 51 void NAME##N##x##W##_##R(NAME##N##x##W##_ctr_t ctr, NAME##N##x##W##_ukey_t ukey, NAME##N##x##W##_ctr_t kactr, unsigned count, CUDAInfo *tp) \ 52 { \ 53 const char *kernelname = PREFIX #NAME #N "x" #W "_" #R; \ 54 NAME##N##x##W##_ctr_t *hC, *dC; \ 55 int n, niterations = numtrials; /* we make niterations + 2 (warmup, overhead) calls to the kernel */ \ 56 double cur_time; \ 57 size_t i; \ 58 const size_t nworkitems = tp->blocks_per_grid * tp->threads_per_block; \ 59 const size_t szC = nworkitems*sizeof(hC[0]); \ 60 \ 61 /* allocate vector of counters in device memory, initialize to zero */ \ 62 CHECKCALL(cudaMalloc(&dC, szC)); \ 63 CHECKCALL(cudaMemset((void *)dC, 0, szC)); \ 64 /* allocate and initialize vector of counters in host memory */ \ 65 CHECKNOTZERO(hC = (NAME##N##x##W##_ctr_t *)malloc(szC)); \ 66 for (i = 0; i < nworkitems; i++) { \ 67 size_t xi; \ 68 for (xi = 0; xi < N; xi++) { \ 69 hC[i].v[xi] = 0; \ 70 } \ 71 } \ 72 unsigned kcount = 0; \ 73 double basetime = 0., dt = 0., mindt = 0.; \ 74 for (n = -2; n < niterations; n++) { \ 75 if (n == -2) { \ 76 if (count == 0) { \ 77 /* try to set a good guess for count */ \ 78 count = (unsigned)(tp->cycles ? tp->cycles * 1e-8 : 10000); \ 79 dprintf(("starting with count = %u\n", count)); \ 80 } \ 81 kcount = count; \ 82 } else if (n == -1) { \ 83 /* use first iteration time to calibrate count to get approximately sec_per_trial */ \ 84 if (count > 1) { \ 85 count = (unsigned)(count * sec_per_trial / dt); \ 86 dprintf(("scaled count = %u\n", count)); \ 87 } \ 88 /* second iteration is to calculate overhead after warmup */ \ 89 kcount = 1; \ 90 } else if (n == 0) { \ 91 int xj; \ 92 /* Check that we got the expected value */ \ 93 for (xj = 0; xj < N; xj++) { \ 94 if (kactr.v[xj] != hC[0].v[xj]) { \ 95 printf("%s mismatch: xj = %d, expected\n", kernelname, xj); \ 96 printline_##NAME##N##x##W##_##R(ukey, ctr, &kactr, 1); \ 97 printf(" but got\n"); \ 98 printline_##NAME##N##x##W##_##R(ukey, ctr, hC, 1); \ 99 if(!debug) exit(1); \ 100 else break; \ 101 } else { \ 102 dprintf(("%s matched word %d\n", kernelname, xj)); \ 103 } \ 104 } \ 105 basetime = dt; \ 106 if (debug||verbose) { \ 107 dprintf(("%s %.3f secs for %lu workitems test on device %d (%s)\n", \ 108 kernelname, basetime, (unsigned long) nworkitems, \ 109 tp->devnum, tp->dev.name)); \ 110 printline_##NAME##N##x##W##_##R(ukey, ctr, hC, (verbose < 2) ? 1 : nworkitems); \ 111 } \ 112 kcount = count + 1; \ 113 } \ 114 dprintf(("launch %s\n", kernelname)); \ 115 (void)timer(&cur_time); \ 116 test_##NAME##N##x##W##_##R<<<tp->blocks_per_grid, tp->threads_per_block>>>(kcount, ctr, ukey, dC); \ 117 dprintf(("synchronize\n")); \ 118 CHECKCALL(cudaDeviceSynchronize()); \ 119 dprintf(("copy results back from device to host\n")); \ 120 CHECKCALL(cudaMemcpy(hC, dC, szC, cudaMemcpyDeviceToHost)); \ 121 dt = timer(&cur_time); \ 122 dprintf(("iteration %d took %.3f secs\n", n, dt)); \ 123 ALLZEROS(hC, nworkitems, N); \ 124 if (n == 0 || dt < mindt) mindt = dt; \ 125 } \ 126 if (count > 1) { \ 127 double tpB = (mindt - basetime) / ( (kcount - 1.) * nworkitems * (N * W / 8.) ); \ 128 printf("%-17s %#5.3g cpB, %#5.3g GB/s on dev %d %u B granularity (best %u in %.3f s - %.6f s)\n", \ 129 kernelname + sizeof(PREFIX) - 1, \ 130 tpB * tp->cycles , 1e-9/tpB, \ 131 tp->devnum, (unsigned)(N*W/8), kcount, mindt, basetime ); \ 132 fflush(stdout); \ 133 } \ 134 /* free host and device memory */ \ 135 free(hC); \ 136 CHECKCALL(cudaFree(dC)); \ 137 } 138 139 #include "util_expandtpl.h" 140 141 int main(int argc, char **argv) 142 { 143 char *cp; 144 unsigned count = 0; 145 CUDAInfo *infop; 146 int keyctroffset = 0; 147 148 progname = argv[0]; 149 if (argc > 3|| (argv[1] && argv[1][0] == '-')) { 150 fprintf(stderr, "Usage: %s [COUNT [DEVICESTRING]]\n", progname); 151 exit(1); 152 } 153 if (argc > 1) 154 count = atoi(argv[1]); 155 if ((cp = getenv("TIME_CUDA_VERBOSE")) != NULL) { 156 verbose = atoi(cp); 157 } 158 if ((cp = getenv("TIME_CUDA_DEBUG")) != NULL) { 159 debug = atoi(cp); 160 } 161 if ((cp = getenv("TIME_CUDA_OFFSET")) != NULL) { 162 keyctroffset = atoi(cp); 163 } 164 if ((cp = getenv("TIME_CUDA_NUMTRIALS")) != NULL) { 165 numtrials = atoi(cp); 166 } 167 if ((cp = getenv("TIME_CUDA_SEC_PER_TRIAL")) != NULL) { 168 sec_per_trial = atof(cp); 169 } 170 infop = cuda_init(argc > 2 ? argv[2] : NULL); 171 /* define macro to initialize and call the kernels */ 172 # include "time_initkeyctr.h" 173 cuda_done(infop); 174 return 0; 175 }