Mac OSx 上的 OpenCL 内核错误

我编写了一些 OpenCL 代码，该代码在 LINUX 上运行良好，但在 Mac OSX 上失败并出现错误。有人可以帮我确定为什么会发生这些情况吗？错误后显示内核代码。我的内核使用 double，所以我在顶部有相应的编译指示。但我不知道为什么错误显示浮点数据类型：

inline float8 __OVERLOAD__ _name(float8 x) { return _default_name(x); } \
                       ^
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:4606:30: note: candidate function
__CLFN_FD_1FD_FAST_RELAX(__fast_relax_log, native_log, __cl_log);
                         ^
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:421:29: 

note: expanded from macro '__CLFN_FD_1FD_FAST_RELAX'
inline float16 __OVERLOAD__ _name(float16 x){ return _default_name(x); }
                        ^
<program source>:206:19: error: call to '__fast_relax_log' is ambiguous
                                    det_zkinin + log((2.0) * 3.14));
              ^~~~~~~~~~~~~~~~~
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:4608:22: 
note: expanded from macro 'log'
#define log(__x) __fast_relax_log(__x)
                 ^~~~~~~~~~~~~~~~
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:4606:30: 
note: candidate function
__CLFN_FD_1FD_FAST_RELAX(__fast_relax_log, native_log, __cl_log);
                         ^
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:416:27: 

note: expanded from macro '__CLFN_FD_1FD_FAST_RELAX'
inline float __OVERLOAD__ _name(float x) { return _default_name(x); } \
                      ^
/System/Library/Frameworks/OpenCL.framework/Versions/A/lib/clang/3.2/include/cl_kernel.h:4606:30 
note: candidate function
__CLFN_FD_1FD_FAST_RELAX(__fast_relax_log, native_log, __cl_log);
                         ^

                       ^

这是内核代码：

#pragma OPENCL EXTENSION cl_khr_fp64: enable

__kernel void ckf_kernel2(int dimx, int aligned_dimx, 
                          int numOfCKF, int aligned_ckf,
                          int iter, 
                          double epsilon,
                          __global double * yrlists, 
                          __global double * zrlists,
                          __global double * rlists,
                          __global double * init_state,
                          __global double * init_var,
                          __global double * sing_j,
                          __global double * covMatrixSum,
                          __global double * cummulative,
                          __global double * temp_var,
                          __global double * x_k_f,
                          __global double * z_k_j,
                          __global double * crossCovMatrixSum,
                          __global double * z_k_f,
                          __global double * innCovMatrixSum,
                          __global double * zk_diff,
                          __global double * reduce_gain_matrix,
                          __global double * llk
    )
{

    int ckf_id = get_global_id(0);

    if( ckf_id < numOfCKF){



        for (int i = 0 ; i < dimx ; i++)
        {
            for (int idx = 0; idx < dimx * 2 ; idx++)
            {
                int column = idx % dimx;
                int mode = (idx >= dimx) ? -1 : 1;
                sing_j[(i * dimx * 2 + idx) * aligned_ckf + ckf_id] = temp_var[(i * dimx + column) * aligned_ckf + ckf_id] * epsilon * mode + init_state[i * aligned_ckf + ckf_id];

            }
        }
        z_k_f[ckf_id] = 0;
        innCovMatrixSum[ckf_id] = 0;
        for (int idx = 0; idx < dimx * 2 ; idx++)
        {
            z_k_j[idx * aligned_ckf + ckf_id] = 0;
            for (int i = 0 ; i < dimx ; i++)
                z_k_j[idx * aligned_ckf + ckf_id] += sing_j[(i * dimx * 2 + idx) * aligned_ckf + ckf_id] * zrlists[iter * aligned_dimx + i ];

            z_k_f[ckf_id] += z_k_j[idx* aligned_ckf + ckf_id] ;
            innCovMatrixSum[ckf_id]  += z_k_j[idx* aligned_ckf + ckf_id] * z_k_j[idx* aligned_ckf + ckf_id];
        }
        z_k_f[ckf_id] = z_k_f[ckf_id]  / (dimx * 2);
        innCovMatrixSum[ckf_id] = innCovMatrixSum[ckf_id] / (dimx * 2);
        innCovMatrixSum[ckf_id] = (innCovMatrixSum[ckf_id] - z_k_f[ckf_id] *z_k_f[ckf_id]) + rlists[ckf_id];

        // calcualte crossCovMatrixSum
        for (int idx = 0; idx < dimx; idx ++)
        {

            crossCovMatrixSum[idx * aligned_ckf + ckf_id] = 0;
            for (int i = 0 ; i < 2 * dimx ; i++)
            {
                crossCovMatrixSum[idx * aligned_ckf + ckf_id] += sing_j[(idx * dimx*2 + i) * aligned_ckf + ckf_id ] * z_k_j[i* aligned_ckf + ckf_id];
            }   
            crossCovMatrixSum[idx * aligned_ckf + ckf_id] = crossCovMatrixSum[idx * aligned_ckf + ckf_id]/ (dimx * 2);
            crossCovMatrixSum[idx * aligned_ckf + ckf_id] = crossCovMatrixSum[idx * aligned_ckf + ckf_id] - x_k_f[idx* aligned_ckf + ckf_id] * z_k_f[ckf_id];

        }

        // calculate zk_diff

        int z_check = (int)yrlists[iter];
        if (z_check == -1)
            zk_diff[ckf_id] = 0;
        else
            zk_diff[ckf_id] = yrlists[iter] - z_k_f[ckf_id];


        // calculate reduce_gain_matrix and (reduce_state_matrix  <==> init_state);
        for (int idx = 0 ; idx < dimx; idx++)
        {
            reduce_gain_matrix[idx * aligned_ckf + ckf_id] =  (crossCovMatrixSum[idx * aligned_ckf + ckf_id] / innCovMatrixSum[ckf_id]);
            init_state[idx * aligned_ckf + ckf_id] =    reduce_gain_matrix[idx * aligned_ckf + ckf_id] * zk_diff[ckf_id] + x_k_f[idx* aligned_ckf + ckf_id];

        }

        for (int idx = 0 ; idx < dimx; idx++)
        {
            init_var[idx * aligned_ckf + ckf_id ] = covMatrixSum[(idx * dimx + idx) * aligned_ckf + ckf_id] - 
                reduce_gain_matrix[idx * aligned_ckf + ckf_id] * innCovMatrixSum[ckf_id] *
                reduce_gain_matrix[idx * aligned_ckf + ckf_id];

        }

        double det_zkinin = zk_diff[ckf_id] * zk_diff[ckf_id] * (1.0f /innCovMatrixSum[ckf_id]);

        if (innCovMatrixSum[ckf_id] <= 0)
            llk[ckf_id] = 0;
        else
            llk[ckf_id] = 0.5 * ((log(innCovMatrixSum[ckf_id])) + 
                                 det_zkinin + log((2.0) * 3.14));

        cummulative[ckf_id] += llk[ckf_id];
    }

}

我怀疑您正在尝试在不支持双精度的集成英特尔 GPU 上运行它。如果我为 Intel HD 4000 编译内核代码，我只能在我自己的 Macbook Pro 上重现您的错误 - 当我针对 CPU 或独立 NVIDIA GPU 时，它编译得很好。

您可以通过查询来检查设备是否支持双精度CL_DEVICE_DOUBLE_FP_CONFIG设备信息参数：

cl_device_fp_config cfg;
clGetDeviceInfo(device, CL_DEVICE_DOUBLE_FP_CONFIG, sizeof(cfg), &cfg, NULL);
printf("Double FP config = %llu\n", cfg);

如果这个函数返回值0，则不支持双精度。这解释了为什么编译器日志只报告float的变种log功能。