Method GetKernelBody

compiler/lib/KernelGen/GeneralIntrinsic/Reduce.cpp:625–704

}

std::string ReduceKernel::GetKernelBody(TContext* context) const {
    auto mode = context->getAttrStr("mode");
    int axis = context->getAttrInt("axis");
    auto input = context->getAttrOprand("operand:0");
    auto src_dtype = input.dtype;
    std::stringstream writer;
    writer << R"(
#include "gi_float.h"
#include "gi_int.h"
    )";
    if (src_dtype == "f16") {
        writer << R"(
#include "gi_float16.h"
static inline gi_float16_t max(gi_float16_t a,gi_float16_t b){return a>b?a:b;}
static inline gi_float16_t min(gi_float16_t a,gi_float16_t b){return a<b?a:b;}    
    )";
        GIMathHelper gi_math;
        if ("MIN" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceMinNanFloat16() << "\n";
        } else if ("MAX" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceMaxNanFloat16() << "\n";
        } else if ("SUM" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceAddFloat16() << "\n";
        } else if ("SUM_SQR" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceAddFloat16() << "\n";
            writer << gi_math.GiMultiplyAddFloat16() << "\n";
        } else if ("MEAN" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceAddFloat16() << "\n";
        } else if ("PRODUCT" == mode) {
            writer << gi_math.FastFp32toFp16() << "\n";
            writer << gi_math.GiReduceMultiplyFloat16() << "\n";
        }
    } else {
        writer << R"(
static inline float max(float a,float b){return a>b?a:b;}
static inline float min(float a,float b){return a<b?a:b;}    
    )";
    }
    writer << GenCommonRet() << " " << GetKernelSignature(context);
    std::string tmp_body = R"({
    const size_t SIMD_WIDTH = ${simd_width};
    const size_t axis = ${axis};
    ${src_specifier}* src = (${src_specifier}*)inputs[0]->ptr;
    ${src_specifier}* dst = (${src_specifier}*)outputs[0]->ptr;
    Layout in_layout = inputs[0]->layout;
    size_t A = 1, B = 0, C = 1;
    for (size_t i = 0; i < axis; ++ i)
        A *= in_layout.dims[i];
    B = in_layout.dims[axis];
    for (size_t i = axis + 1; i < in_layout.nr_dim; ++ i){
        C *= in_layout.dims[i];
    }
    if(C == 1){