Method GetCVKernelBody

compiler/lib/KernelGen/Arm/ArmCommon/Resize.cpp:104–778

}

std::string ResizeKernel::GetCVKernelBody(TContext* context) const {
    auto kernel_sig = GetCVKernelSignature(context);
    std::string body_temp = R"(
        #include <math.h>
        #include <stdalign.h>
        #include <string.h>
        #include "marm_neon.h"
        #include "tinycv_c.h"
        #define rep(i, n) for (int i = 0; i < (n); ++i)
        static inline uint8_t output_converter(float x){
            x = fmin(255.0f, fmax(0.0f, x));
            return (uint8_t) roundf(x);
        }

        static inline void get_nearest_linear_coord(float scale, int size, int idx, float* ah0, int* ih0, float* ah1, int* ih1){
            if (size == 1) {
                *ah0 = 1.f;
                *ih0 = 0;
                *ah1 = 0.f;
                *ih1 = 0;
            }

            float alpha = (idx + 0.5f) / scale - 0.5f;
            int origin_idx = (int)(floorf(alpha));
            alpha -= origin_idx;

            if (origin_idx < 0) {
                origin_idx = 0;
                alpha = 0;
            } else if (origin_idx + 1 >= size) {
                origin_idx = size - 2;
                alpha = 1;
            }

            *ah0 = 1 - alpha;
            *ih0 = origin_idx;
            *ah1 = alpha;
            *ih1 = origin_idx + 1;
        }
#ifdef RESIZE_INT_IMPLEMENT
        #define SCALE 11
        static inline void calc_cache_8uc3_1(const uint8_t* src, const int src_h_stride, const uint8_t* dst, const int OW,
                       const int* tabsx,
                       const int* tabsy,
                       const int* tabrx,
                       const int* tabry, int dx,
                       int* cache0, int* cache1) {
            (void)tabrx;
            const uint8_t* psrc1 = src + (tabsx[dx] + 1) * src_h_stride;

            int dy = 0, dy3 = 0;

            for (; dy < OW; ++dy, dy3 += 3) {
                const uint8_t* pcsrc10 = psrc1 + (tabsy[dy] + 0) * 3;
                const uint8_t* pcsrc11 = psrc1 + (tabsy[dy] + 1) * 3;
                int ry = tabry[dy];
                int iry = (1 << SCALE) - ry;
                cache1[dy3 + 0] = pcsrc11[0] * ry + pcsrc10[0] * iry;