// Auto-generated file. Do not edit!
//   Template: src/f32-igemm/neon-shuffle.c.in
//   Generator: tools/xngen
//
// Copyright 2019 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 <arm_neon.h>

#include <xnnpack/igemm.h>


void xnn_f32_igemm_ukernel_1x8s4__neon(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const float**restrict a,
    const float*restrict w,
    float*restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const float* zero,
    const union xnn_f32_minmax_params params[restrict static 1])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  float* c0 = c;

  do {
    float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
    float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;

    size_t p = ks;
    do {
      const float* restrict a0 = a[0];
      assert(a0 != NULL);
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const float*) ((uintptr_t) a0 + a_offset);
      }
      a += 1;

      size_t k = kc;
      while (k >= 4 * sizeof(float)) {
        float32x4_t va0 = vld1q_f32(a0); a0 += 4;


        const float32x4_t vb0123c0 = vld1q_f32(w + 0);
        const float32x4_t vb4567c0 = vld1q_f32(w + 4);

        vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
        vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);

        va0 = vextq_f32(va0, va0, 1);

        const float32x4_t vb0123c1 = vld1q_f32(w + 8);
        const float32x4_t vb4567c1 = vld1q_f32(w + 12);

        vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
        vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);

        va0 = vextq_f32(va0, va0, 1);

        const float32x4_t vb0123c2 = vld1q_f32(w + 16);
        const float32x4_t vb4567c2 = vld1q_f32(w + 20);

        vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
        vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);

        va0 = vextq_f32(va0, va0, 1);

        const float32x4_t vb0123c3 = vld1q_f32(w + 24);
        const float32x4_t vb4567c3 = vld1q_f32(w + 28);

        vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
        vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);


        w += 32;
        k -= 4 * sizeof(float);
      }
      if XNN_UNLIKELY(k != 0) {
        do {
          const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;

          const float32x4_t vb0123 = vld1q_f32(w); w += 4;
          const float32x4_t vb4567 = vld1q_f32(w); w += 4;

          vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
          vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);

          k -= sizeof(float);
        } while (k != 0);
      }

      p -= 1 * sizeof(void*);
    } while (p != 0);

    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    vacc0x0123 = vminq_f32(vacc0x0123, vmax);
    vacc0x4567 = vminq_f32(vacc0x4567, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
    vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);

    if XNN_LIKELY(nc >= 8) {
      vst1q_f32(c0, vacc0x0123);
      vst1q_f32(c0 + 4, vacc0x4567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= 8;
    } else {
      if (nc & 4) {
        vst1q_f32(c0, vacc0x0123); c0 += 4;

        vacc0x0123 = vacc0x4567;
      }
      float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
      if (nc & 2) {
        vst1_f32(c0, vacc0x01); c0 += 2;

        vacc0x01 = vget_high_f32(vacc0x0123);
      }
      if (nc & 1) {
        vst1_lane_f32(c0, vacc0x01, 0);
      }

      nc = 0;
    }
  } while (nc != 0);
}