// Auto-generated file. Do not edit!
//   Template: src/x8-packw/scalar.c.in
//   Generator: tools/xngen
//
// Copyright 2023 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.


#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include "xnnpack/math.h"
#include "xnnpack/packw.h"
#include "xnnpack/unaligned.h"

void xnn_x8_packw_gemm_goi_ukernel_x2__scalar_u4(
  size_t g,
  size_t nc,
  size_t kc,
  size_t nr,
  size_t kr,
  size_t sr,
  const int8_t* weights,
  const uint32_t* bias,
  const void* scale,
  int8_t* packed_weights,
  size_t extra_bytes,
  const void* params)
{
  assert(g != 0);
  assert(nc != 0);
  assert(kc != 0);
  assert(nr == 2);   // This kernel is for NR=2
  assert(kr == 1);
  assert(sr == 1);
  assert(weights != NULL);
  assert(packed_weights != NULL);

  int8_t* out = (int8_t*) packed_weights;
  const uint32_t* b = (const uint32_t*) bias;

  do {
    // NC main loop multiple of 2
    const int8_t* w0 = (const int8_t*) weights;
    size_t n = nc;
    for (;n >= 2; n -= 2) {
      if XNN_LIKELY(b != NULL) {
        unaligned_store_s32(out + 0 * sizeof(int32_t), b[0]);
        unaligned_store_s32(out + 1 * sizeof(int32_t), b[1]);
        b += 2;
      } else {
        unaligned_store_s32(out + 0 * sizeof(int32_t), 0);
        unaligned_store_s32(out + 1 * sizeof(int32_t), 0);
      }
      out += 2 * sizeof(uint32_t);

      const int8_t* w1 = w0 + kc;

      // KC main loop multiple of 2x4
      size_t k = kc;
      for (; k >= 4; k -= 4) {
        const int8_t v00 = w0[0];
        const int8_t v01 = w0[1];
        const int8_t v02 = w0[2];
        const int8_t v03 = w0[3];
        w0 += 4;
        const int8_t v10 = w1[0];
        const int8_t v11 = w1[1];
        const int8_t v12 = w1[2];
        const int8_t v13 = w1[3];
        w1 += 4;
        out[0] = v00;
        out[1] = v10;
        out[2] = v01;
        out[3] = v11;
        out[4] = v02;
        out[5] = v12;
        out[6] = v03;
        out[7] = v13;
        out += 8;
      }

      // KC remainder
      for (; k != 0; --k) {
        const int8_t v0 = *w0++;
        out[0] = v0;
        const int8_t v1 = *w1++;
        out[1] = v1;
        out += 2;
      }
      out = (int8_t*) ((uintptr_t) out + extra_bytes);
      w0 = w1;
    }

    // NC remainder (1..1)
    if XNN_UNLIKELY(n != 0) {
      if XNN_LIKELY(b != NULL) {
        size_t nb = n;
        do {
          unaligned_store_s32(out, *b++);
          out += sizeof(uint32_t);
        } while (--nb != 0);
      } else {
        size_t nb = n;
        do {
          unaligned_store_s32(out, 0);
          out += sizeof(uint32_t);
        } while (--nb != 0);
      }
      out += (2 - n) * sizeof(uint32_t);


      // KC main loop multiple of 2x4
      size_t k = kc;
      for (; k >= 4; k -= 4) {
        const int8_t v00 = w0[0];
        const int8_t v01 = w0[1];
        const int8_t v02 = w0[2];
        const int8_t v03 = w0[3];
        w0 += 4;
        out[0] = v00;
        out[2] = v01;
        out[4] = v02;
        out[6] = v03;
        out += 8;
      }

      // KC remainder of 1..3
      for (; k != 0; --k) {
        const int8_t v0 = *w0++;
        out[0] = v0;
        out += 2;
      }
      out = (int8_t*) ((uintptr_t) out + extra_bytes);
    }
    weights += nc * kc;
  } while (--g != 0);
}