Function cactus_max_all_f16

cactus/kernel/kernel_reduce.cpp:201–222

}

__fp16 cactus_max_all_f16(const __fp16* data, size_t num_elements) {
    return CactusThreading::parallel_reduce(
        num_elements, CactusThreading::Thresholds::ALL_REDUCE,
        [&](size_t start, size_t end) -> __fp16 {
            const size_t vec_end = start + simd_align(end - start);
            float16x8_t acc = vdupq_n_f16(static_cast<__fp16>(-65504.0f));

            for (size_t i = start; i < vec_end; i += SIMD_F16_WIDTH) {
                acc = vmaxq_f16(acc, vld1q_f16(&data[i]));
            }

            __fp16 result = static_cast<__fp16>(-65504.0f);
            __fp16 arr[8];
            vst1q_f16(arr, acc);
            for (int j = 0; j < 8; j++) result = std::max(result, arr[j]);
            for (size_t i = vec_end; i < end; ++i) result = std::max(result, data[i]);
            return result;
        },
        static_cast<__fp16>(-65504.0f),
        [](__fp16 a, __fp16 b) { return std::max(a, b); }
    );
}

void cactus_max_axis_f16(const __fp16* input, __fp16* output, size_t outer_size, size_t axis_size, size_t inner_size) {
    axis_reduce_f16_impl(input, output, outer_size, axis_size, inner_size,