三星Galaxy S7上OpenCL half_与native_函数编译失败求助

Let's break down why you're hitting CL_BUILD_FAILURE when using half_* or native_* functions, and how to fix it:

1. First, Get the Exact Build Error #

You haven't shared the full build log, which is critical for pinpointing the issue. Add this code right after your build call to see what's going wrong:

cl_int buildStatus = program_RANSACEllipses.getBuildInfo<CL_PROGRAM_BUILD_STATUS>(devices[0]);
if (buildStatus != CL_SUCCESS) {
    std::string buildLog = program_RANSACEllipses.getBuildInfo<CL_PROGRAM_BUILD_LOG>(devices[0]);
    std::cerr << "OpenCL Build Log:\n" << buildLog << std::endl;
}

This will tell you if the error is "undefined function" (driver doesn't support the half-native functions) or an extension-related issue.

2. Mali-T880's OpenCL FP16 Limitation #

While your device supports cl_khr_fp16 and OpenCL 1.2, Mali-T880's driver has a key limitation: it does not implement native_* functions for half types. The native_sin, native_cos, etc., are only available for float inputs. This is a hardware/driver constraint, not an issue with your code or headers.

3. Fixes to Try #

Option 1: Work Around Missing Half-Native Functions #

Convert your half values to float, use the native_* float functions, then convert back to half. This adds minimal overhead and works reliably:

#pragma OPENCL EXTENSION cl_khr_fp16 : enable
half distancePointToEllipse(short2 point, half2 ecenter, half phi, half a, half b) {
    half2 t = convert_half2(point) - ecenter;
    half angle_half = atan2(t.y, t.x) - phi;
    
    // Convert to float for native functions
    float angle_float = convert_float(angle_half);
    float cos1_float = native_cos(angle_float);
    float sin1_float = native_sin(angle_float);
    
    // Convert back to half
    half cos1 = convert_half(cos1_float);
    half sin1 = convert_half(sin1_float);
    
    half tx = a * a * cos1 * cos1;
    half ty = b * b * sin1 * sin1;
    half c = native_sqrt(convert_float(tx + ty)); // Same trick for sqrt
    half d = hypot(t.x, t.y);
    return fabs(d - c);
}

Option 2: Ensure Extension is Properly Enabled #

Even though you have the #pragma in your kernel, some Mali drivers require explicitly enabling the extension via build flags. Update your build call to:

program_RANSACEllipses.build("-cl-std=CL1.2 -cl-ext=+cl_khr_fp16");

This forces the compiler to recognize the cl_khr_fp16 extension, which can resolve subtle declaration issues.

Option 3: Stick to Standard FP16 Functions #

You're already using sincos (a standard OpenCL function that works with half on Mali-T880). If performance isn't significantly impacted, keep using standard functions like sqrt instead of native_sqrt—they're more portable and guaranteed to work with your driver.

4. Verify Device Capabilities #

Double-check that your device actually supports the functions you're trying to use. Run this code to list all built-in functions supported by the GPU:

size_t funcListSize;
clGetDeviceInfo(devices[0], CL_DEVICE_BUILT_IN_KERNELS, 0, nullptr, &funcListSize);
std::vector<char> funcList(funcListSize);
clGetDeviceInfo(devices[0], CL_DEVICE_BUILT_IN_KERNELS, funcListSize, funcList.data(), nullptr);
std::cout << "Built-in functions:\n" << funcList.data() << std::endl;

You won't see native_sin(half) or similar entries here, confirming the driver doesn't implement them.

内容的提问来源于stack exchange，提问作者mompes