Method GetKernelBody

compiler/lib/KernelGen/GeneralIntrinsic/Typecvt.cpp:227–293

}  // namespace

std::string TypecvtKernel::GetKernelBody(TContext* context) const {
    std::stringstream ss;
    auto src_dtype_str =
            SymbolHelper::gen_valid_dtype(context->getAttrOprand("operand:0").dtype);
    auto dst_dtype_str =
            SymbolHelper::gen_valid_dtype(context->getAttrOprand("operand:1").dtype);
    std::string src_specifier = Utils::cvt_dtype_specifier(src_dtype_str);
    std::string dst_specifier = Utils::cvt_dtype_specifier(dst_dtype_str);
    ss << R"(
    #include "gi_float.h"
    #include "gi_int.h"
    )";
    GIMathHelper gi_math;
    if (Utils::is_float_dtype(src_dtype_str, 16) ||
        Utils::is_float_dtype(dst_dtype_str, 16)) {
        ss << "#include \"gi_float16.h\"\n";
        ss << gi_math.FastFp32toFp16() << "\n";
        ss << gi_math.FastFp16toFp32() << "\n";
    }
    ss << GenCommonRet() << " " << GetKernelSignature(context);
    std::string body_temp = R"({
    ${init_declare_str}
    const Tensor* src_tensor = inputs[0];
    const Tensor* dst_tensor = outputs[0];
    ${src_specifier}* src = (${src_specifier}*)(src_tensor->ptr);
    ${dst_specifier}* dst = (${dst_specifier}*)(dst_tensor->ptr);
    TINYNN_ASSERT(src);
    TINYNN_ASSERT(dst);
    
    const Layout src_layout = src_tensor->layout;
    const Layout dst_layout = dst_tensor->layout;
    float src_scale = src_tensor->dtype.param.scale;
    float dst_scale = dst_tensor->dtype.param.scale;

    size_t nr_elem = 1;
    for (int i = 0; i < src_layout.nr_dim; ++i) {
        nr_elem *= src_layout.dims[i];
    }
    ${gen_scale}
    scale = src_scale/dst_scale;
    vscale = GiBroadcastFloat32(scale);
    size_t idx = 0;
    
    for(; idx + SIMD_WIDTH <= nr_elem; idx += SIMD_WIDTH){
        ${gen_cvt}
        src += SIMD_WIDTH;
        dst += SIMD_WIDTH;
    }

    for(;idx < nr_elem;++idx){
        ${gen_cvt_remain}
        ++src;
        ++dst;
    }
    return TinyNN_SUCCESS;
})";

    ss << StringTemplate::StringTemplateArgs()