gpufetch/src/cuda/cuda.cpp

#include <cuda_runtime.h>
#include <cstring>
#include <cstdlib>
#include <cstdio>

#include "cuda.hpp"
#include "uarch.hpp"
#include "pci.hpp"
#include "gpufetch_helper_cuda.hpp"
#include "../common/global.hpp"
#include "../common/uarch.hpp"

bool print_gpu_cuda(struct gpu_info* gpu) {
  char* cc = get_str_cc(gpu->arch);
  printf("%s (Compute Capability %s)\n", gpu->name, cc);
  free(cc);

  return true;
}

struct cache* get_cache_info(cudaDeviceProp prop) {
  struct cache* cach = (struct cache*) emalloc(sizeof(struct cache));

  cach->L2 = (struct cach*) emalloc(sizeof(struct cach));
  cach->L2->size = prop.l2CacheSize;
  cach->L2->num_caches = 1;
  cach->L2->exists = true;

  return cach;
}

int get_tensor_cores(struct uarch* arch, int sm, int major) {
  if(major == 7) {
    // TU116 does not have tensor cores!
    // https://www.anandtech.com/show/13973/nvidia-gtx-1660-ti-review-feat-evga-xc-gaming/2
    if (is_chip_TU116(arch))
      return 0;
    return sm * 8;
  }
  else if(major == 8) return sm * 4;
  else return 0;
}

struct topology_c* get_topology_info(struct uarch* arch, cudaDeviceProp prop) {
  struct topology_c* topo = (struct topology_c*) emalloc(sizeof(struct topology_c));

  topo->streaming_mp = prop.multiProcessorCount;
  topo->cores_per_mp = _ConvertSMVer2Cores(prop.major, prop.minor);
  topo->cuda_cores = topo->streaming_mp * topo->cores_per_mp;
  topo->tensor_cores = get_tensor_cores(arch, topo->streaming_mp, prop.major);

  return topo;
}

int32_t guess_clock_multipilier(struct gpu_info* gpu, struct memory* mem) {
  // Guess clock multiplier
  int32_t clk_mul = 1;

  int32_t clk8 = abs((mem->freq/8) - gpu->freq);
  int32_t clk4 = abs((mem->freq/4) - gpu->freq);
  int32_t clk2 = abs((mem->freq/2) - gpu->freq);
  int32_t clk1 = abs((mem->freq/1) - gpu->freq);

  int32_t min = mem->freq;
  if(clkm_possible_for_uarch(8, gpu->arch) && min > clk8) { clk_mul = 8; min = clk8; }
  if(clkm_possible_for_uarch(4, gpu->arch) && min > clk4) { clk_mul = 4; min = clk4; }
  if(clkm_possible_for_uarch(2, gpu->arch) && min > clk2) { clk_mul = 2; min = clk2; }
  if(clkm_possible_for_uarch(1, gpu->arch) && min > clk1) { clk_mul = 1; min = clk1; }

  return clk_mul;
}

struct memory* get_memory_info(struct gpu_info* gpu, cudaDeviceProp prop) {
  struct memory* mem = (struct memory*) emalloc(sizeof(struct memory));

  mem->size_bytes = (unsigned long long) prop.totalGlobalMem;
  mem->freq = prop.memoryClockRate * 0.001f;
  mem->bus_width = prop.memoryBusWidth;
  mem->clk_mul = guess_clock_multipilier(gpu, mem);
  mem->type = guess_memtype_from_cmul_and_uarch(mem->clk_mul, gpu->arch);

  // Fix frequency returned from CUDA to show real frequency
  mem->freq = mem->freq  / mem->clk_mul;

  return mem;
}

// Compute peak performance when using CUDA cores
int64_t get_peak_performance_cuda(struct gpu_info* gpu) {
  return gpu->freq * 1000000 * gpu->topo_c->cuda_cores * 2;
}

// Compute peak performance when using tensor cores
int64_t get_peak_performance_tcu(cudaDeviceProp prop, struct gpu_info* gpu) {
  // Volta / Turing tensor cores performs 4x4x4 FP16 matrix multiplication
  // Ampere tensor cores performs 8x4x8 FP16 matrix multiplicacion
  if(prop.major == 7) return gpu->freq * 1000000 * 4 * 4 * 4  * 2 * gpu->topo_c->tensor_cores;
  else if(prop.major == 8) return gpu->freq * 1000000 * 8 * 4 * 8 * 2 * gpu->topo_c->tensor_cores;
  else return 0;
}

struct gpu_info* get_gpu_info_cuda(struct pci_dev *devices, int gpu_idx) {
  struct gpu_info* gpu = (struct gpu_info*) emalloc(sizeof(struct gpu_info));
  gpu->pci = NULL;
  gpu->idx = gpu_idx;

  if(gpu->idx < 0) {
    printErr("GPU index must be equal or greater than zero");
    return NULL;
  }

  if(gpu_idx == 0) {
    printf("%s", CUDA_DRIVER_START_WARNING);
    fflush(stdout);
  }

  int num_gpus = -1;
  cudaError_t err = cudaSuccess;
  err = cudaGetDeviceCount(&num_gpus);

  if(gpu_idx == 0) {
    printf("\r%*c\r", (int) strlen(CUDA_DRIVER_START_WARNING), ' ');
    fflush(stdout);
  }

  if(err != cudaSuccess) {
    printErr("%s: %s", cudaGetErrorName(err), cudaGetErrorString(err));
    return NULL;
  }

  if(num_gpus <= 0) {
    printErr("No CUDA capable devices found!");
    return NULL;
  }

  if(gpu->idx+1 > num_gpus) {
    // Master is trying to query an invalid GPU
    return NULL;
  }

  cudaDeviceProp deviceProp;
  if ((err = cudaGetDeviceProperties(&deviceProp, gpu->idx)) != cudaSuccess) {
    printErr("%s: %s", cudaGetErrorName(err), cudaGetErrorString(err));
    return NULL;
  }

  gpu->freq = deviceProp.clockRate * 1e-3f;
  gpu->vendor = GPU_VENDOR_NVIDIA;
  gpu->name = (char *) emalloc(sizeof(char) * (strlen(deviceProp.name) + 1));
  strcpy(gpu->name, deviceProp.name);

  if((gpu->pci = get_pci_from_pciutils(devices, PCI_VENDOR_ID_NVIDIA, gpu_idx)) == NULL) {
    printErr("Unable to find a valid device for vendor id 0x%.4X using pciutils", PCI_VENDOR_ID_NVIDIA);
    return NULL;
  }
  gpu->arch = get_uarch_from_cuda(gpu);
  gpu->cach = get_cache_info(deviceProp);
  gpu->mem = get_memory_info(gpu, deviceProp);
  gpu->topo_c = get_topology_info(gpu->arch, deviceProp);
  gpu->peak_performance = get_peak_performance_cuda(gpu);
  gpu->peak_performance_tcu = get_peak_performance_tcu(deviceProp, gpu);

  return gpu;
}

char* get_str_sm(struct gpu_info* gpu) {
  return get_str_generic(gpu->topo_c->streaming_mp);
}

char* get_str_cores_sm(struct gpu_info* gpu) {
  return get_str_generic(gpu->topo_c->cores_per_mp);
}

char* get_str_cuda_cores(struct gpu_info* gpu) {
  return get_str_generic(gpu->topo_c->cuda_cores);
}

char* get_str_tensor_cores(struct gpu_info* gpu) {
  return get_str_generic(gpu->topo_c->tensor_cores);
}