CUDA Samples 中 1_Utilities 文件夹里包含了一些实用工具和小型示例程序,它们通常用于支持和演示其他 CUDA 示例程序的功能。
deviceQueryDrv和deviceQuery 类似，主要是是查询系统上安装的 CUDA 设备的属性。区别是deviceQuery 使用了cuda runtime API, deviceQueryDrv使用了CUDA driverAPI。这个程序可以帮助开发人员了解系统上可用的 CUDA 设备的具体信息,以便编写更高效的 CUDA 应用程序。
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <cuda.h>
#include <helper_cuda_drvapi.h>
// Program main
int main(int argc, char **argv) {
CUdevice dev;
int major = 0, minor = 0;
int deviceCount = 0;
char deviceName[256];
printf("%s Starting...\n\n", argv[0]);
// note your project will need to link with cuda.lib files on windows
printf("CUDA Device Query (Driver API) statically linked version \n");
checkCudaErrors(cuInit(0));
checkCudaErrors(cuDeviceGetCount(&deviceCount));
// This function call returns 0 if there are no CUDA capable devices.
if (deviceCount == 0) {
printf("There are no available device(s) that support CUDA\n");
} else {
printf("Detected %d CUDA Capable device(s)\n", deviceCount);
}
for (dev = 0; dev < deviceCount; ++dev) {
checkCudaErrors(cuDeviceGetAttribute(
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, dev));
checkCudaErrors(cuDeviceGetAttribute(
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, dev));
checkCudaErrors(cuDeviceGetName(deviceName, 256, dev));
printf("\nDevice %d: \"%s\"\n", dev, deviceName);
int driverVersion = 0;
checkCudaErrors(cuDriverGetVersion(&driverVersion));
printf(" CUDA Driver Version: %d.%d\n",
driverVersion / 1000, (driverVersion % 100) / 10);
printf(" CUDA Capability Major/Minor version number: %d.%d\n", major,
minor);
size_t totalGlobalMem;
checkCudaErrors(cuDeviceTotalMem(&totalGlobalMem, dev));
char msg[256];
SPRINTF(msg,
" Total amount of global memory: %.0f MBytes "
"(%llu bytes)\n",
(float)totalGlobalMem / 1048576.0f,
(unsigned long long)totalGlobalMem);
printf("%s", msg);
int multiProcessorCount;
getCudaAttribute<int>(&multiProcessorCount,
CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev);
printf(" (%2d) Multiprocessors, (%3d) CUDA Cores/MP: %d CUDA Cores\n",
multiProcessorCount, _ConvertSMVer2CoresDRV(major, minor),
_ConvertSMVer2CoresDRV(major, minor) * multiProcessorCount);
int clockRate;
getCudaAttribute<int>(&clockRate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, dev);
printf(
" GPU Max Clock rate: %.0f MHz (%0.2f "
"GHz)\n",
clockRate * 1e-3f, clockRate * 1e-6f);
int memoryClock;
getCudaAttribute<int>(&memoryClock, CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE,
dev);
printf(" Memory Clock rate: %.0f Mhz\n",
memoryClock * 1e-3f);
int memBusWidth;
getCudaAttribute<int>(&memBusWidth,
CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH, dev);
printf(" Memory Bus Width: %d-bit\n",
memBusWidth);
int L2CacheSize;
getCudaAttribute<int>(&L2CacheSize, CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE, dev);
if (L2CacheSize) {
printf(" L2 Cache Size: %d bytes\n",
L2CacheSize);
}
int maxTex1D, maxTex2D[2], maxTex3D[3];
getCudaAttribute<int>(&maxTex1D,
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH, dev);
getCudaAttribute<int>(&maxTex2D[0],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH, dev);
getCudaAttribute<int>(&maxTex2D[1],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT, dev);
getCudaAttribute<int>(&maxTex3D[0],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH, dev);
getCudaAttribute<int>(&maxTex3D[1],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT, dev);
getCudaAttribute<int>(&maxTex3D[2],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH, dev);
printf(
" Max Texture Dimension Sizes 1D=(%d) 2D=(%d, %d) "
"3D=(%d, %d, %d)\n",
maxTex1D, maxTex2D[0], maxTex2D[1], maxTex3D[0], maxTex3D[1],
maxTex3D[2]);
int maxTex1DLayered[2];
getCudaAttribute<int>(&maxTex1DLayered[0],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH,
dev);
getCudaAttribute<int>(&maxTex1DLayered[1],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS,
dev);
printf(
" Maximum Layered 1D Texture Size, (num) layers 1D=(%d), %d layers\n",
maxTex1DLayered[0], maxTex1DLayered[1]);
int maxTex2DLayered[3];
getCudaAttribute<int>(&maxTex2DLayered[0],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH,
dev);
getCudaAttribute<int>(&maxTex2DLayered[1],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT,
dev);
getCudaAttribute<int>(&maxTex2DLayered[2],
CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS,
dev);
printf(
" Maximum Layered 2D Texture Size, (num) layers 2D=(%d, %d), %d "
"layers\n",
maxTex2DLayered[0], maxTex2DLayered[1], maxTex2DLayered[2]);
int totalConstantMemory;
getCudaAttribute<int>(&totalConstantMemory,
CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY, dev);
printf(" Total amount of constant memory: %u bytes\n",
totalConstantMemory);
int sharedMemPerBlock;
getCudaAttribute<int>(&sharedMemPerBlock,
CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK, dev);
printf(" Total amount of shared memory per block: %u bytes\n",
sharedMemPerBlock);
int regsPerBlock;
getCudaAttribute<int>(®sPerBlock,
CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK, dev);
printf(" Total number of registers available per block: %d\n",
regsPerBlock);
int warpSize;
getCudaAttribute<int>(&warpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, dev);
printf(" Warp size: %d\n", warpSize);
int maxThreadsPerMultiProcessor;
getCudaAttribute<int>(&maxThreadsPerMultiProcessor,
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
dev);
printf(" Maximum number of threads per multiprocessor: %d\n",
maxThreadsPerMultiProcessor);
int maxThreadsPerBlock;
getCudaAttribute<int>(&maxThreadsPerBlock,
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, dev);
printf(" Maximum number of threads per block: %d\n",
maxThreadsPerBlock);
int blockDim[3];
getCudaAttribute<int>(&blockDim[0], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X,
dev);
getCudaAttribute<int>(&blockDim[1], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y,
dev);
getCudaAttribute<int>(&blockDim[2], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z,
dev);
printf(" Max dimension size of a thread block (x,y,z): (%d, %d, %d)\n",
blockDim[0], blockDim[1], blockDim[2]);
int gridDim[3];
getCudaAttribute<int>(&gridDim[0], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, dev);
getCudaAttribute<int>(&gridDim[1], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y, dev);
getCudaAttribute<int>(&gridDim[2], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z, dev);
printf(" Max dimension size of a grid size (x,y,z): (%d, %d, %d)\n",
gridDim[0], gridDim[1], gridDim[2]);
int textureAlign;
getCudaAttribute<int>(&textureAlign, CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT,
dev);
printf(" Texture alignment: %u bytes\n",
textureAlign);
int memPitch;
getCudaAttribute<int>(&memPitch, CU_DEVICE_ATTRIBUTE_MAX_PITCH, dev);
printf(" Maximum memory pitch: %u bytes\n",
memPitch);
int gpuOverlap;
getCudaAttribute<int>(&gpuOverlap, CU_DEVICE_ATTRIBUTE_GPU_OVERLAP, dev);
int asyncEngineCount;
getCudaAttribute<int>(&asyncEngineCount,
CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, dev);
printf(
" Concurrent copy and kernel execution: %s with %d copy "
"engine(s)\n",
(gpuOverlap ? "Yes" : "No"), asyncEngineCount);
int kernelExecTimeoutEnabled;
getCudaAttribute<int>(&kernelExecTimeoutEnabled,
CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, dev);
printf(" Run time limit on kernels: %s\n",
kernelExecTimeoutEnabled ? "Yes" : "No");
int integrated;
getCudaAttribute<int>(&integrated, CU_DEVICE_ATTRIBUTE_INTEGRATED, dev);
printf(" Integrated GPU sharing Host Memory: %s\n",
integrated ? "Yes" : "No");
int canMapHostMemory;
getCudaAttribute<int>(&canMapHostMemory,
CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, dev);
printf(" Support host page-locked memory mapping: %s\n",
canMapHostMemory ? "Yes" : "No");
int concurrentKernels;
getCudaAttribute<int>(&concurrentKernels,
CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS, dev);
printf(" Concurrent kernel execution: %s\n",
concurrentKernels ? "Yes" : "No");
int surfaceAlignment;
getCudaAttribute<int>(&surfaceAlignment,
CU_DEVICE_ATTRIBUTE_SURFACE_ALIGNMENT, dev);
printf(" Alignment requirement for Surfaces: %s\n",
surfaceAlignment ? "Yes" : "No");
int eccEnabled;
getCudaAttribute<int>(&eccEnabled, CU_DEVICE_ATTRIBUTE_ECC_ENABLED, dev);
printf(" Device has ECC support: %s\n",
eccEnabled ? "Enabled" : "Disabled");
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
int tccDriver;
getCudaAttribute<int>(&tccDriver, CU_DEVICE_ATTRIBUTE_TCC_DRIVER, dev);
printf(" CUDA Device Driver Mode (TCC or WDDM): %s\n",
tccDriver ? "TCC (Tesla Compute Cluster Driver)"
: "WDDM (Windows Display Driver Model)");
#endif
int unifiedAddressing;
getCudaAttribute<int>(&unifiedAddressing,
CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING, dev);
printf(" Device supports Unified Addressing (UVA): %s\n",
unifiedAddressing ? "Yes" : "No");
int managedMemory;
getCudaAttribute<int>(&managedMemory, CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY,
dev);
printf(" Device supports Managed Memory: %s\n",
managedMemory ? "Yes" : "No");
int computePreemption;
getCudaAttribute<int>(&computePreemption,
CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED,
dev);
printf(" Device supports Compute Preemption: %s\n",
computePreemption ? "Yes" : "No");
int cooperativeLaunch;
getCudaAttribute<int>(&cooperativeLaunch,
CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH, dev);
printf(" Supports Cooperative Kernel Launch: %s\n",
cooperativeLaunch ? "Yes" : "No");
int cooperativeMultiDevLaunch;
getCudaAttribute<int>(&cooperativeMultiDevLaunch,
CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH,
dev);
printf(" Supports MultiDevice Co-op Kernel Launch: %s\n",
cooperativeMultiDevLaunch ? "Yes" : "No");
int pciDomainID, pciBusID, pciDeviceID;
getCudaAttribute<int>(&pciDomainID, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, dev);
getCudaAttribute<int>(&pciBusID, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, dev);
getCudaAttribute<int>(&pciDeviceID, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, dev);
printf(" Device PCI Domain ID / Bus ID / location ID: %d / %d / %d\n",
pciDomainID, pciBusID, pciDeviceID);
const char *sComputeMode[] = {
"Default (multiple host threads can use ::cudaSetDevice() with device "
"simultaneously)",
"Exclusive (only one host thread in one process is able to use "
"::cudaSetDevice() with this device)",
"Prohibited (no host thread can use ::cudaSetDevice() with this "
"device)",
"Exclusive Process (many threads in one process is able to use "
"::cudaSetDevice() with this device)",
"Unknown", NULL};
int computeMode;
getCudaAttribute<int>(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, dev);
printf(" Compute Mode:\n");
printf(" < %s >\n", sComputeMode[computeMode]);
}
// If there are 2 or more GPUs, query to determine whether RDMA is supported
if (deviceCount >= 2) {
int gpuid[64]; // we want to find the first two GPUs that can support P2P
int gpu_p2p_count = 0;
int tccDriver = 0;
for (int i = 0; i < deviceCount; i++) {
checkCudaErrors(cuDeviceGetAttribute(
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, i));
checkCudaErrors(cuDeviceGetAttribute(
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, i));
getCudaAttribute<int>(&tccDriver, CU_DEVICE_ATTRIBUTE_TCC_DRIVER, i);
// Only boards based on Fermi or later can support P2P
if ((major >= 2)
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
// on Windows (64-bit), the Tesla Compute Cluster driver for windows
// must be enabled to support this
&& tccDriver
#endif
) {
// This is an array of P2P capable GPUs
gpuid[gpu_p2p_count++] = i;
}
}
// Show all the combinations of support P2P GPUs
int can_access_peer;
char deviceName0[256], deviceName1[256];
if (gpu_p2p_count >= 2) {
for (int i = 0; i < gpu_p2p_count; i++) {
for (int j = 0; j < gpu_p2p_count; j++) {
if (gpuid[i] == gpuid[j]) {
continue;
}
checkCudaErrors(
cuDeviceCanAccessPeer(&can_access_peer, gpuid[i], gpuid[j]));
checkCudaErrors(cuDeviceGetName(deviceName0, 256, gpuid[i]));
checkCudaErrors(cuDeviceGetName(deviceName1, 256, gpuid[j]));
printf(
"> Peer-to-Peer (P2P) access from %s (GPU%d) -> %s (GPU%d) : "
"%s\n",
deviceName0, gpuid[i], deviceName1, gpuid[j],
can_access_peer ? "Yes" : "No");
}
}
}
}
printf("Result = PASS\n");
exit(EXIT_SUCCESS);
}
代码详解：
程序会依次输出每个设备的详细信息,包括:
-
设备数量和设备属性:
- 设备数量
- 每个设备的名称
- 每个设备的计算能力主版本号和次版本号
- 每个设备的总内存大小
-
设备功能属性:
- 每个设备最大的线程数、线程块数和网格大小
- 每个设备支持的最大纹理维度和最大纹理数组大小
- 每个设备支持的最大常量内存大小和寄存器数量
- 每个设备支持的 ECC 错误检查和纠正
-
设备内存相关属性:
- 每个设备的global内存带宽
- 每个设备的shared内存每个块的大小
- 每个设备的constant内存大小
-
设备计算能力相关属性:
- 每个设备的计算能力主/次版本号
- 每个设备的 warp 大小
- 每个设备的最大线程维度
- 每个设备的最大网格维度
CUDA 相关API解读
该Sample使用的CUDA API: cuInit, cuDeviceGetCount, cuDeviceComputeCapability, cuDriverGetVersion, cuDeviceTotalMem, cuDeviceGetAttribute
这些API属于CUDA driver API。
1.cuInit
需要先使用 cuInit() 函数初始化driverAPI,然后才能调用driver API 中的任何函数。
接下来必须创建一个与特定设备关联的 CUDA 上下文(context),并将其设置为当前调用线程的活跃上下文。
- 首先调用
cuInit()函数来初始化驱动 API。 - 然后创建一个与特定设备关联的 CUDA 上下文。
- 将创建的 CUDA 上下文设置为当前调用线程的活跃上下文。
- 之后就可以开始使用驱动 API 中的函数了。
这个初始化和上下文设置步骤是在通过驱动 API 与 CUDA 设备进行交互之前必须完成的前提条件。
在 CUDA 上下文中,内核(kernels)需要由device代码显式地以 PTX 或二进制对象的形式加载,因此,用 C++ 编写的内核必须单独编译成 PTX 或二进制对象。内核的启动使用 API 入口点。
任何想要在未来设备架构上运行的应用程序都必须加载 PTX 代码,而不是二进制代码。这是因为二进制代码是特定于架构的,因此与未来的架构不兼容,而 PTX 代码则是在加载时由设备驱动程序编译成二进制代码的。
以下是使用driver API 编写的内核示例的主机代码:
int main()
{
int N = ...;
size_t size = N * sizeof(float);
// Allocate input vectors h_A and h_B in host memory
float* h_A = (float*)malloc(size);
float* h_B = (float*)malloc(size);
// Initialize input vectors
...
// Initialize
cuInit(0);
// Get number of devices supporting CUDA
int deviceCount = 0;
cuDeviceGetCount(&deviceCount);
if (deviceCount == 0) {
printf("There is no device supporting CUDA.\n");
exit (0);
}
// Get handle for device 0
CUdevice cuDevice;
cuDeviceGet(&cuDevice, 0);
// Create context
CUcontext cuContext;
cuCtxCreate(&cuContext, 0, cuDevice);
// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, "VecAdd.ptx");
// Allocate vectors in device memory
CUdeviceptr d_A;
cuMemAlloc(&d_A, size);
CUdeviceptr d_B;
cuMemAlloc(&d_B, size);
CUdeviceptr d_C;
cuMemAlloc(&d_C, size);
// Copy vectors from host memory to device memory
cuMemcpyHtoD(d_A, h_A, size);
cuMemcpyHtoD(d_B, h_B, size);
// Get function handle from module
CUfunction vecAdd;
cuModuleGetFunction(&vecAdd, cuModule, "VecAdd");
// Invoke kernel
int threadsPerBlock = 256;
int blocksPerGrid =
(N + threadsPerBlock - 1) / threadsPerBlock;
void* args[] = { &d_A, &d_B, &d_C, &N };
cuLaunchKernel(vecAdd,
blocksPerGrid, 1, 1, threadsPerBlock, 1, 1,
0, 0, args, 0);
...
}2. cuDeviceGetCount
该API用于获取系统中可用的 CUDA 设备的数量。
3. cuDeviceComputeCapability
该函数用于获取指定 CUDA 设备的计算能力。
4.cuDriverGetVersion
该函数用于获取当前系统上安装的 CUDA driver的版本号。
5.cuDeviceTotalMem
该函数用于获取指定 CUDA 设备的总内存大小
6. cuDeviceGetAttribue
该函数用于获取指定 CUDA 设备的属性值。
除此之外，代码中大量调用getCudaAttribute函数，该函数是被包装在helper_cuda_drvapi.h 里的模板函数 ，实际调用的是cuDeviceGetAttribue API。
运行结果：
CUDA Device Query (Driver API) statically linked version
Detected 1 CUDA Capable device(s)
Device 0: "NVIDIA GeForce RTX 4080"
CUDA Driver Version: 12.5
CUDA Capability Major/Minor version number: 8.9
Total amount of global memory: 16376 MBytes (17170956288 bytes)
(76) Multiprocessors, (128) CUDA Cores/MP: 9728 CUDA Cores
GPU Max Clock rate: 2505 MHz (2.50 GHz)
Memory Clock rate: 11201 Mhz
Memory Bus Width: 256-bit
L2 Cache Size: 67108864 bytes
Max Texture Dimension Sizes 1D=(131072) 2D=(131072, 65536) 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 1536
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Texture alignment: 512 bytes
Maximum memory pitch: 2147483647 bytes
Concurrent copy and kernel execution: Yes with 1 copy engine(s)
Run time limit on kernels: Yes
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Concurrent kernel execution: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
CUDA Device Driver Mode (TCC or WDDM): WDDM (Windows Display Driver Model)
Device supports Unified Addressing (UVA): Yes
Device supports Managed Memory: Yes
Device supports Compute Preemption: Yes
Supports Cooperative Kernel Launch: Yes
Supports MultiDevice Co-op Kernel Launch: No
Device PCI Domain ID / Bus ID / location ID: 0 / 1 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >
Result = PASS
996
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
