time_opencl.c (7558B)
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 * OpenCL 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 #define R123_USE_AES_NI 0 /* never use this for OpenCL */ 39 40 #include "util_opencl.h" 41 42 #include "Random123/philox.h" 43 #include "Random123/threefry.h" 44 45 #include "time_misc.h" 46 #include "util_print.h" 47 48 #define TEST_TPL(NAME, N, W, R) \ 49 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, UCLInfo *tp) \ 50 { \ 51 const char *kernelname = PREFIX #NAME #N "x" #W "_" #R; \ 52 NAME##N##x##W##_ctr_t *hC; \ 53 int n, niterations = numtrials; /* we make niterations + 2 (warmup, overhead) calls to the kernel */ \ 54 int narg; \ 55 double cur_time; \ 56 size_t i; \ 57 cl_int err; \ 58 cl_mem dC; \ 59 cl_kernel kern; \ 60 \ 61 /* create handle to kernel in program */ \ 62 dprintf(("%s kernel\n", kernelname)); \ 63 CHECKERR(kern = clCreateKernel(tp->prog, kernelname, &err)); \ 64 const size_t nworkitems = tp->wgsize * tp->cores; \ 65 const size_t szC = nworkitems*sizeof(hC[0]); \ 66 /* allocate and initialize vector of counters in host memory */ \ 67 CHECKNOTZERO(hC = (NAME##N##x##W##_ctr_t *) malloc(szC)); \ 68 for (i = 0; i < nworkitems; i++) { \ 69 int xi; \ 70 for (xi = 0; xi < N; xi++) \ 71 hC[i].v[xi] = 0; \ 72 } \ 73 /* allocate vector of counters in device memory, initialize from current host memory */ \ 74 CHECKERR(dC = clCreateBuffer(tp->ctx, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, szC, hC, &err)); \ 75 cl_uint kcount = 0; \ 76 double basetime = 0., dt = 0., mindt = 0.; \ 77 /* first two iterations are for warmup & baseline */ \ 78 for (n = -2; n < niterations; n++) { \ 79 if (n == -2) { \ 80 if (count == 0) { \ 81 /* try to set a good guess for count */ \ 82 count = (unsigned)(tp->cycles ? tp->cycles * 1e-8 : 10000); \ 83 dprintf(("starting with count = %u\n", count)); \ 84 } \ 85 kcount = count; \ 86 } else if (n == -1) { \ 87 /* use first iteration time to calibrate count to get approximately sec_per_trial */ \ 88 if (count > 1) { \ 89 count = (unsigned)(count * sec_per_trial / dt); \ 90 dprintf(("scaled count = %u\n", count)); \ 91 } \ 92 /* second iteration is to calculate overhead after warmup */ \ 93 kcount = 1; \ 94 } else if (n == 0) { \ 95 int xj; \ 96 /* Check that we got the expected value */ \ 97 for (xj = 0; xj < N; xj++) { \ 98 if (kactr.v[xj] != hC[0].v[xj]) { \ 99 printf("%s mismatch: xj = %d, expected\n", kernelname, xj); \ 100 printline_##NAME##N##x##W##_##R(ukey, ctr, &kactr, 1); \ 101 printf(" but got\n"); \ 102 printline_##NAME##N##x##W##_##R(ukey, ctr, hC, 1); \ 103 if(!debug) exit(1); \ 104 else break; \ 105 } else { \ 106 dprintf(("%s matched word %d\n", kernelname, xj)); \ 107 } \ 108 } \ 109 basetime = dt; \ 110 if (debug||verbose) { \ 111 dprintf(("%s %.3f secs for %lu workitems test on device %s\n", \ 112 kernelname, basetime, (unsigned long) nworkitems, \ 113 tp->devname)); \ 114 printline_##NAME##N##x##W##_##R(ukey, ctr, hC, (verbose < 2) ? 1 : nworkitems); \ 115 } \ 116 kcount = count + 1; \ 117 } \ 118 (void)timer(&cur_time); \ 119 dprintf(("setup arguments to kernel function %s\n", kernelname)); \ 120 narg = 0; \ 121 CHECK(clSetKernelArg(kern, narg, sizeof(kcount), (void *)&kcount)); \ 122 narg++; \ 123 CHECK(clSetKernelArg(kern, narg, sizeof(ctr), (void *)&ctr)); \ 124 narg++; \ 125 CHECK(clSetKernelArg(kern, narg, sizeof(ukey), (void *)&ukey)); \ 126 narg++; \ 127 CHECK(clSetKernelArg(kern, narg, sizeof(cl_mem), (void *)&dC)); \ 128 dprintf(("queue kernel for execution on device %s\n", tp->devname)); \ 129 CHECK(clEnqueueNDRangeKernel(tp->cmdq, kern, 1, 0, &nworkitems, &tp->wgsize, 0, 0, 0)); \ 130 CHECK(clFlush(tp->cmdq)); \ 131 dprintf(("copy results from device memory to host memory\n")); \ 132 CHECK(clEnqueueReadBuffer(tp->cmdq, dC, CL_FALSE, 0, szC, hC, 0, 0, 0)); \ 133 CHECK(clFlush(tp->cmdq)); \ 134 dprintf(("synchronize\n")); \ 135 CHECK(clFinish(tp->cmdq)); \ 136 dt = timer(&cur_time); \ 137 dprintf(("iteration %d took %.3f secs\n", n, dt)); \ 138 ALLZEROS(hC, nworkitems, N); \ 139 if (n == 0 || dt < mindt) mindt = dt; \ 140 } \ 141 if (count > 1) { \ 142 double tpB = (mindt - basetime) / ( (count - 1.) * nworkitems * (N * W / 8.) ); \ 143 printf("%-17s %#5.3g cpB, %#5.3g GB/s %u B granularity (best %u in %.3f s - %.6f s)\n", \ 144 kernelname + sizeof(PREFIX) - 1, \ 145 tpB * tp->cycles , 1e-9/tpB, \ 146 (unsigned)(N*W/8), count, mindt, basetime ); \ 147 fflush(stdout); \ 148 } \ 149 /* free host and device memory */ \ 150 free(hC); \ 151 CHECK(clReleaseMemObject(dC)); \ 152 /* free the kernel */ \ 153 CHECK(clReleaseKernel(kern)); \ 154 } 155 156 #include "util_expandtpl.h" 157 158 /* Include the preprocessed source code, stashed in a literal string */ 159 const char *opencl_src = 160 #include "time_opencl_kernel.i" 161 ; 162 163 int main(int argc, char **argv) 164 { 165 const char *cp; 166 unsigned count = 0; 167 UCLInfo *infop; 168 int keyctroffset = 0; 169 170 progname = argv[0]; 171 if (argc > 3|| (argv[1] && argv[1][0] == '-')) { 172 fprintf(stderr, "Usage: %s [COUNT [DEVICESTRING]]\n", progname); 173 exit(1); 174 } 175 if (argc > 1) 176 count = atoi(argv[1]); 177 if ((cp = getenv("TIME_OPENCL_VERBOSE")) != NULL) { 178 verbose = atoi(cp); 179 } 180 if ((cp = getenv("TIME_OPENCL_DEBUG")) != NULL) { 181 debug = atoi(cp); 182 } 183 if ((cp = getenv("TIME_OPENCL_OFFSET")) != NULL) { 184 keyctroffset = atoi(cp); 185 } 186 if ((cp = getenv("TIME_OPENCL_NUMTRIALS")) != NULL) { 187 numtrials = atoi(cp); 188 } 189 if ((cp = getenv("TIME_OPENCL_SEC_PER_TRIAL")) != NULL) { 190 sec_per_trial = atof(cp); 191 } 192 if ((cp = getenv("TIME_OPENCL_BUILD_OPTIONS")) == NULL) { 193 cp = ""; 194 } 195 infop = opencl_init(argc > 2 ? argv[2] : NULL, opencl_src, cp); 196 /* define macro to initialize and call the kernels */ 197 # include "time_initkeyctr.h" 198 opencl_done(infop); 199 return 0; 200 }