// Auto-generated file. Do not edit!
//   Template: src/f32-igemm/hvx-broadcast.c.in
//   Generator: tools/xngen
//
// Copyright 2024 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 <hexagon_types.h>
#include <hexagon_protos.h>
#include <hvx_hexagon_protos.h>

#include "xnnpack/igemm.h"
#include "xnnpack/intrinsics-polyfill.h"

void xnn_f32_igemm_minmax_ukernel_1x64__hvx_broadcast(
    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 XNN_MIN_ELEMENTS(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 {
    HVX_Vector vacc0x0 = *((HVX_Vector *)(w));
    HVX_Vector vacc0x1 = *((HVX_Vector *)(w + 32));
    w += 64;

    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;
      do {
        const HVX_Vector vb0 = *((HVX_Vector *)(w));
        const HVX_Vector vb1 = *((HVX_Vector *)(w + 32));
        w += 64;

        const HVX_Vector va0 =  Q6_V_vsplat_R(*(uint32_t *)a0);
        a0 += 1;

        vacc0x0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(va0, vb0), vacc0x0));
        vacc0x1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(va0, vb1), vacc0x1));
        k -= sizeof(float);
      } while (k != 0);
      p -= 1 * sizeof(void*);
    } while (p != 0);

    const HVX_Vector vmin = Q6_V_vsplat_R(params->scalar.min);
    vacc0x0 = Q6_Vw_vmax_VwVw(vmin, vacc0x0);
    vacc0x1 = Q6_Vw_vmax_VwVw(vmin, vacc0x1);

    const HVX_Vector vmax = Q6_V_vsplat_R(params->scalar.max);
    vacc0x0 = Q6_Vw_vmin_VwVw(vmax, vacc0x0);
    vacc0x1 = Q6_Vw_vmin_VwVw(vmax, vacc0x1);

    if XNN_LIKELY(nc >= 64) {
      *((HVX_UVector *)(c0)) = vacc0x0;
      *((HVX_UVector *)(c0 + 32)) = vacc0x1;
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= 64;
    } else {
      if (nc & 32) {
        *((HVX_UVector *)c0) = vacc0x0;

        vacc0x0 = vacc0x1;

        c0 += 32;
        nc ^= 32;
      }
      vstu_variable_scalar((char*)c0, nc*sizeof(float), vacc0x0);

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