1

请帮助...在 opencl 程序中，我收到这些错误。我没有“oclUtils.h”头文件。S0 我没有将它包含在我的程序中。这就是我遇到这个问题的原因。我也在发送我的内核。如果这是原因，我如何在不使用 oclUtils.h 的情况下使用原子函数？

*"__kernel void BLAS_susdot_kernel(__global float *x,__global int *index,__global float        *y,const int n,__global float* dot_p)                    \n" \
"{                                                                                      \n" \
"    int block_x = get_group_id(0);                                                                     \n" \
"    int thread_x = get_local_id(0);                                                                    \n" \
"    int i = get_global_id(0);                                                                          \n" \
"    float dot_value, old = *dot_p;                                                                     \n" \
"    int warp_thread_id = i & (32-1);                                                                   \n" \
"    __local float tmp[512];                                                                                    \n" \
"    __local float share_dot_p;                                                                         \n" \
"    if(thread_x == 0)                          \n" \
"       share_dot_p = 0.0;                              \n" \
"    if (i < n)                                 \n" \
"   {                                        \n" \
"        tmp[i]= x[i]*y[index[i]];           \n" \
"        if(warp_thread_id <16 && (i+16)< n) \n" \
"            tmp[i]+=tmp[i+16];              \n" \
"        if(warp_thread_id <8 && (i+8)< n )  \n" \
"            tmp[i]+=tmp[i+8];               \n" \
"        if(warp_thread_id <4 && (i+4)< n )  \n" \
"            tmp[i]+=tmp[i+4];               \n" \
"        if(warp_thread_id <2 && (i+2)< n)   \n" \
"            tmp[i]+=tmp[i+2];               \n" \
"        if(warp_thread_id==0 && (i+1)< n)   \n" \
"        {                                   \n" \
"            tmp[i]+=tmp[i+1];               \n" \
"            do                              \n" \
"            {                               \n" \
"                dot_value = old;            \n" \
"                old = convert_int( atomic_cmpxchg((volatile __global unsigned int*)&share_dot_p, convert_int(dot_value), convert_int(tmp[i] + dot_value)));  \n" \
"            } while (dot_value != old);     \n" \
"            //share_dot_p +=tmp[i];         \n" \
"        }                                   \n" \
"        if(thread_x==0)                  \n" \
"        {                                   \n" \
"                                            \n" \
"            do                              \n" \
"            {                               \n" \
"                dot_value = old;            \n" \
"                old = convert_int(atomic_cmpxchg((volatile __global unsigned int *) dot_p, convert_int(dot_value), convert_int(share_dot_p+dot_value)));  \n" \
"            } while(dot_value != old);                                                 \n" \
"               // *dot_p += share_dot_p;                                               \n" \
"                                                                                       \n" \
"        }                                                                              \n" \
"   }                                                                                   \n" \
"}                                                                                      \n" \
"\n";*

我得到的错误：

Build Program Info: ptxas application ptx input, line 160; error   : Label expected for argument 0 of instruction 'call'
ptxas application ptx input, line 160; error   : Call target not recognized
ptxas application ptx input, line 160; error   : Function 'atomic_cmpxchg' not declared in this scope
ptxas application ptx input, line 160; error   : Call target not recognized
ptxas application ptx input, line 185; error   : Label expected for argument 0 of instruction 'call'
ptxas application ptx input, line 185; error   : Call target not recognized
ptxas application ptx input, line 185; error   : Function 'atomic_cmpxchg' not declared in this scope
ptxas application ptx input, line 185; error   : Call target not recognized
ptxas application ptx input, line 161; error   : Unknown symbol 'atomic_cmpxchg'
ptxas application ptx input, line 186; error   : Unknown symbol 'atomic_cmpxchg'
ptxas fatal   : Ptx assembly aborted due to errors
error   : Ptx compilation failed: gpu='sm_13', device code='cuModuleLoadDataEx_4'
: Considering profile 'compute_13' for gpu='sm_13' in 'cuModuleLoadDataEx_4'
: Retrieving binary for 'cuModuleLoadDataEx_4', for gpu='sm_13', usage mode='  '
: Considering profile 'compute_13' for gpu='sm_13' in 'cuModuleLoadDataEx_4'
: Control flags for 'cuModuleLoadDataEx_4' disable search path
: Ptx binary found for 'cuModuleLoadDataEx_4', architecture='compute_13'
: Ptx compilation for 'cuModuleLoadDataEx_4', for gpu='sm_13', ocg options='  '
4

1 回答 1

2

由于您使用atomic_cmpxchg的是 OpenCL 中可选扩展的一部分，因此您需要检查您的设备是否支持它，然后在内核代码中启用它：

  1. 检查cl_khr_global_int32_base_atomics是否在返回的扩展名中列出clGetDeviceInfo(..., CL_DEVICE_EXTENSIONS, ...)

  2. 将以下内容添加到内核代码的顶部：

#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable

请注意，如果您使用带有 __local 或 64 位操作数的原子函数，您可能需要启用其他扩展：

#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
#pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
#pragma OPENCL EXTENSION cl_khr_int64_extended_atomics : enable

此外，请确保您使用的是 OpenCL 1.1 或更高版本，因为在 OpenCL 1.0 中原子函数的名称不同。例如，如果您使用 NVidia 的nvcc实用程序离线编译内核，请确保您指定了-arch命令行开关。

这些特定错误与 oclUtils.h 无关。这是一个 NVidia 头文件，仅当您在主机（而非内核）代码中调用 ocl* 函数时才需要它。

于 2012-11-11T19:56:07.980 回答