src/amalgam/gen/fp16arith.c

// Copyright 2021 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 <string.h>

#include <arm_fp16.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>
#include <xnnpack/vbinary.h>


void xnn_f16_vdiv_minmax_ukernel__fp16arith_x2(
    size_t batch,
    const void* restrict input_a,
    const void* restrict input_b,
    void* restrict output,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float16_t) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const float16_t* a = (const float16_t*) input_a;
  const float16_t* b = (const float16_t*) input_b;
  float16_t* o = (float16_t*) output;

  float16_t vy_min, vy_max;
  memcpy(&vy_min, &params->fp16arith.min, sizeof(vy_min));
  memcpy(&vy_max, &params->fp16arith.max, sizeof(vy_max));

  for (; batch >= 2 * sizeof(float16_t); batch -= 2 * sizeof(float16_t)) {
    const float16_t va0 = *a++;
    const float16_t va1 = *a++;

    const float16_t vb0 = *b++;
    const float16_t vb1 = *b++;

    float16_t vacc0 = vdivh_f16(va0, vb0);
    float16_t vacc1 = vdivh_f16(va1, vb1);


    vacc0 = vmaxnmh_f16(vacc0, vy_min);
    vacc1 = vmaxnmh_f16(vacc1, vy_min);

    vacc0 = vminnmh_f16(vacc0, vy_max);
    vacc1 = vminnmh_f16(vacc1, vy_max);

    *o++ = vacc0;
    *o++ = vacc1;
  }
  if XNN_UNLIKELY(batch != 0) {
    const float16_t va = *a;
    const float16_t vb = *b;
    float16_t vacc = vdivh_f16(va, vb);
    vacc = vmaxnmh_f16(vacc, vy_min);
    vacc = vminnmh_f16(vacc, vy_max);
    *o = vacc;
  }
}

void xnn_f16_vdivc_minmax_ukernel__fp16arith_x2(
    size_t batch,
    const void* restrict input_a,
    const void* restrict input_b,
    void* restrict output,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float16_t) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const float16_t* a = (const float16_t*) input_a;
  const float16_t* b = (const float16_t*) input_b;
  float16_t* o = (float16_t*) output;

  float16_t vy_min, vy_max;
  memcpy(&vy_min, &params->fp16arith.min, sizeof(vy_min));
  memcpy(&vy_max, &params->fp16arith.max, sizeof(vy_max));

  const float16_t vb = *b;
  for (; batch >= 2 * sizeof(float16_t); batch -= 2 * sizeof(float16_t)) {
    float16_t vacc0 = a[0];
    float16_t vacc1 = a[1];
    a += 2;

    vacc0 = vdivh_f16(vacc0, vb);
    vacc1 = vdivh_f16(vacc1, vb);


    vacc0 = vmaxnmh_f16(vacc0, vy_min);
    vacc1 = vmaxnmh_f16(vacc1, vy_min);

    vacc0 = vminnmh_f16(vacc0, vy_max);
    vacc1 = vminnmh_f16(vacc1, vy_max);

    o[0] = vacc0;
    o[1] = vacc1;
    o += 2;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16_t vacc = *a;
    vacc = vdivh_f16(vacc, vb);
    vacc = vmaxnmh_f16(vacc, vy_min);
    vacc = vminnmh_f16(vacc, vy_max);
    *o = vacc;
  }
}

void xnn_f16_vrdivc_minmax_ukernel__fp16arith_x2(
    size_t batch,
    const void* restrict input_a,
    const void* restrict input_b,
    void* restrict output,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float16_t) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const float16_t* a = (const float16_t*) input_a;
  const float16_t* b = (const float16_t*) input_b;
  float16_t* o = (float16_t*) output;

  float16_t vy_min, vy_max;
  memcpy(&vy_min, &params->fp16arith.min, sizeof(vy_min));
  memcpy(&vy_max, &params->fp16arith.max, sizeof(vy_max));

  const float16_t vb = *b;
  for (; batch >= 2 * sizeof(float16_t); batch -= 2 * sizeof(float16_t)) {
    float16_t vacc0 = a[0];
    float16_t vacc1 = a[1];
    a += 2;

    vacc0 = vdivh_f16(vb, vacc0);
    vacc1 = vdivh_f16(vb, vacc1);


    vacc0 = vmaxnmh_f16(vacc0, vy_min);
    vacc1 = vmaxnmh_f16(vacc1, vy_min);

    vacc0 = vminnmh_f16(vacc0, vy_max);
    vacc1 = vminnmh_f16(vacc1, vy_max);

    o[0] = vacc0;
    o[1] = vacc1;
    o += 2;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16_t vacc = *a;
    vacc = vdivh_f16(vb, vacc);
    vacc = vmaxnmh_f16(vacc, vy_min);
    vacc = vminnmh_f16(vacc, vy_max);
    *o = vacc;
  }
}