random123

Counter-based Random Number Generators
git clone git://git.meso-star.com/random123.git
Log | Files | Refs | README | LICENSE

util_cuda.h (5926B)


      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 #ifndef UTIL_CUDA_H__
     33 #define UTIL_CUDA_H__
     34 
     35 #include "util.h"
     36 
     37 #include <cuda.h>
     38 #include <cuda_runtime_api.h>
     39 
     40 // utility macros to check return codes and complain/exit on failure
     41 #define CHECKLAST(MSG) 	do { cudaError_t e = cudaGetLastError(); if (e != cudaSuccess) {fprintf(stderr, "%s:%d: CUDA Error: %s: %s\n", __FILE__, __LINE__, (MSG), cudaGetErrorString(e)); exit(1); }} while(0)
     42 #define CHECKCALL(RET)	do { cudaError_t e = (RET); if (e != cudaSuccess) { fprintf(stderr, "%s:%d: CUDA Error: %s\n", __FILE__, __LINE__, cudaGetErrorString(e)); exit(1); } } while(0)
     43 
     44 typedef struct cuda_info {
     45     int devnum, cores, blocks_per_grid, threads_per_block;
     46     double cycles;
     47     struct cudaDeviceProp dev;
     48 } CUDAInfo;
     49 
     50 // If devstr is none, chooses device with most cores.
     51 static CUDAInfo *cuda_init(const char *devstr)
     52 {
     53     CUDAInfo *tp;
     54     int i, ndev, cores, devcores, devdev;
     55     double cycles;
     56     CHECKNOTZERO(tp = (CUDAInfo *) malloc(sizeof(CUDAInfo)));
     57     CHECKCALL( cudaGetDeviceCount(&ndev) );
     58     devcores = 0;
     59     devdev = -1;
     60     if (devstr == NULL) {
     61 	devstr = getenv("R123EXAMPLE_ENVCONF_CUDA_DEVICE");
     62 	if (devstr && devstr[0] >= '0' && devstr[0] <= '9' && devstr[1] == '\0') {
     63 	    devdev = devstr[0]-'0';
     64 	}
     65     }
     66     for (i = 0; i < ndev; i++) {
     67 	struct cudaDeviceProp cu;
     68 	CHECKCALL( cudaGetDeviceProperties (&cu, i) );
     69 	// Number of cores is not available from a query, have to hardwire
     70 	// some knowledge here, from web articles about the various generations
     71 	// SM or SMX, might also find this info in
     72 	// CUDA SDK $CUDA_SAMPLES_DIR/common/inc/helper_cuda.h
     73 	// or https://github.com/NVIDIA/nvidia-docker/blob/master/tools/src/cuda/cuda.go
     74 	cores = cu.multiProcessorCount;
     75 	if (cu.major == 1 && cu.minor >= 0 && cu.minor <= 3) {
     76 	    // 1.0 (G80, G92, aka GTX880, Tesla [CSD]870) to 1.3 (GT200, aka GTX280, Tesla [CS]10xx) have 8 cores per MP
     77 	    cores *= 8;
     78 	} else if (cu.major == 2 && cu.minor == 0) {
     79 	    // 2.0 (G100, aka GTX480, Tesla/Fermi [CSM]20[567]0, and GF110, aka GTX580, M2090)
     80 	    cores *= 32;
     81 	} else if (cu.major == 2 && cu.minor == 1) {
     82 	    // 2.1 (GF104, GF114, GF116 aka GTX [45][56]0)
     83 	    cores *= 48;
     84 	} else if (cu.major == 3) {
     85 	    // 3.0 (Kepler GK104 aka GTX 680), 3.2 (TK1), 3.5 (GK11x, GK20x), 3.7 (GK21x)
     86 	    cores *= 192;
     87 	} else if (cu.major == 5) {
     88 	    // 5.0 (Maxwell GM10x), 5.2 (GM20x), 5.3 (TX1)
     89 	    cores *= 128;
     90 	} else if (cu.major == 6 && cu.minor == 0) {
     91 	    // 6.0 (Pascal P100)
     92 	    cores *= 64;
     93 	} else if (cu.major == 6) {
     94 	    // 6.1 (Pascal 10xx, Titan Xp, P40), 6.2 (Drive PX2 and Tegra)
     95 	    cores *= 128;
     96 	} else if (cu.major == 7) {
     97 	    // 7.[05] (Volta and Turing RTX 20[678]0, Titan RTX, Quadro RTX), 7.2 (Xavier Jetson)
     98 	    cores *= 64;
     99 	} else if (cu.major == 8 && cu.minor == 0) {
    100 	    // 8.0 (Ampere A100)
    101 	    cores *= 64;
    102 	} else if (cu.major == 8 && cu.minor == 6) {
    103 	    // 8.6 (Ampere RTX 30[56789]0, A[23456]000, A45000)
    104 	    cores *= 128;
    105 	} else {
    106 	    int coremultguess = 384;
    107 	    cores *= coremultguess;
    108 	    fprintf(stderr, "WARNING: Unknown number of cores per MP for this device: assuming %d, so cpb calculation will be wrong and choice of blocks/grid might be suboptimal\n", coremultguess);
    109 	}
    110 	/* clockrate is in KHz */
    111 	cycles = 1e3 * cu.clockRate * cores;
    112 	printf("  %d: maj %d min %d %s%s ( %d units @ %g MHz ECC=%d %d cores %g Gcycles/s)\n",
    113 	   i, cu.major, cu.minor, cu.name, cu.integrated ? " integrated" : "",
    114 	   cu.multiProcessorCount, cu.clockRate*1e-3, cu.ECCEnabled, cores, cycles*1e-9);
    115 	if ((devstr && strstr(cu.name, devstr) == NULL )||
    116 	    devdev >= 0 && i != devdev) {
    117 	    dprintf(("skipping device %s\n", cu.name));
    118 	    continue;
    119 	}
    120 	if (cores > devcores) {
    121 	    devcores = cores;
    122 	    tp->devnum = i;
    123 	    tp->cores = cores;
    124 	    tp->cycles = cycles;
    125 	    tp->dev = cu;
    126 	}
    127     }
    128     if (devcores == 0) {
    129 	fprintf(stderr, "could not find specified device\n");
    130 	exit(1);
    131     }
    132     tp->blocks_per_grid = tp->cores; /* seems like a good guess */
    133     tp->threads_per_block = tp->dev.warpSize * 2;
    134     printf("Using CUDA device %d, %d cores, %g cycles, will try %d blocks/grid %d threads/block\n",
    135 	   tp->devnum, tp->cores, tp->cycles, tp->blocks_per_grid, tp->threads_per_block);
    136     CHECKCALL(cudaSetDevice(tp->devnum));
    137     dprintf(("cuda_init done\n"));
    138     return tp;
    139 }
    140 
    141 static void cuda_done(CUDAInfo *tp)
    142 {
    143     dprintf(("cuda_done\n"));
    144     free(tp);
    145 }
    146 
    147 #endif /* UTIL_CUDA_H__ */