MR Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/qnnpack/bench/q8gemm.cc lines 219–234

template <uint32_t MR, uint32_t NR, uint32_t NP, uint32_t KR>
class Q8GEMM_L1 : public Q8GEMM {
 public:
  inline Q8GEMM_L1() : Q8GEMM(MR, NR, NP, KR) {
    cpuinfo_initialize();
    const size_t l1d_size = cpuinfo_get_l1d_cache(0)->size;
    const size_t l1d_reserve = 512;
    kc_ = ((l1d_size - l1d_reserve) / sizeof(uint8_t) - mr() * nr()) /
        (mr() + nr());
    if (kr() != 1) {
      kc_ = kc_ / kr() * kr();
    } else {
      kc_ = kc_ / nr() * nr();
    }
  }
};