xref: /aosp_15_r20/external/XNNPACK/src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f32-igemm/MRx2c4-wasmsimd.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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 <wasm_simd128.h>

#include <xnnpack/igemm.h>


void xnn_f32_igemm_minmax_ukernel_4x2c4__wasmsimd_x86(
    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)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= 4);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(ks != 0);
  assert(ks % (4 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  float* c0 = c;
  float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    c1 = c0;
  }
  float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    c2 = c1;
  }
  float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
  if XNN_UNPREDICTABLE(mr != 4) {
    c3 = c2;
  }

  const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
  const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
  do {
    v128_t vacc0x0c4 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, w[0]);
    v128_t vacc0x1c4 = wasm_f32x4_replace_lane(vacc0x0c4, 0, w[1]);
    v128_t vacc1x0c4 = vacc0x0c4;
    v128_t vacc1x1c4 = vacc0x1c4;
    v128_t vacc2x0c4 = vacc0x0c4;
    v128_t vacc2x1c4 = vacc0x1c4;
    v128_t vacc3x0c4 = vacc0x0c4;
    v128_t vacc3x1c4 = vacc0x1c4;
    w += 2;

    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);
      }
      const float* restrict a1 = a[1];
      assert(a1 != NULL);
      if XNN_UNPREDICTABLE(a1 != zero) {
        a1 = (const float*) ((uintptr_t) a1 + a_offset);
      }
      const float* restrict a2 = a[2];
      assert(a2 != NULL);
      if XNN_UNPREDICTABLE(a2 != zero) {
        a2 = (const float*) ((uintptr_t) a2 + a_offset);
      }
      const float* restrict a3 = a[3];
      assert(a3 != NULL);
      if XNN_UNPREDICTABLE(a3 != zero) {
        a3 = (const float*) ((uintptr_t) a3 + a_offset);
      }
      a += 4;

      size_t k = kc;
      for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
        const v128_t va0 = wasm_v128_load(a0);
        a0 += 4;
        const v128_t va1 = wasm_v128_load(a1);
        a1 += 4;
        const v128_t va2 = wasm_v128_load(a2);
        a2 += 4;
        const v128_t va3 = wasm_v128_load(a3);
        a3 += 4;

        const v128_t vb0 = wasm_v128_load(w);
        const v128_t vb1 = wasm_v128_load(w + 4);
        w += 8;

        vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(va0, vb0));
        vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(va0, vb1));
        vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(va1, vb0));
        vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(va1, vb1));
        vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(va2, vb0));
        vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(va2, vb1));
        vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(va3, vb0));
        vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(va3, vb1));
      }
      if XNN_UNLIKELY(k != 0) {
        const v128_t va0 = wasm_v128_load(a0);
        const v128_t va1 = wasm_v128_load(a1);
        const v128_t va2 = wasm_v128_load(a2);
        const v128_t va3 = wasm_v128_load(a3);

        const v128_t vb0 = wasm_v128_load(w);
        const v128_t vb1 = wasm_v128_load(w + 4);
        w += 8;

        const v128_t vzero = wasm_f32x4_const_splat(0.0f);
        const v128_t vmask0 = wasm_f32x4_eq(vb0, vzero);
        const v128_t vmask1 = wasm_f32x4_eq(vb1, vzero);

        vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask0), vb0));
        vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask1), vb1));
        vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask0), vb0));
        vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask1), vb1));
        vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask0), vb0));
        vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask1), vb1));
        vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask0), vb0));
        vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask1), vb1));
      }
      p -= 4 * sizeof(void*);
    } while (p != 0);

    const v128_t vacc0x01c2 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 0, 4, 1, 5),
      wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 2, 6, 3, 7));
    const v128_t vacc1x01c2 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 0, 4, 1, 5),
      wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 2, 6, 3, 7));
    const v128_t vacc2x01c2 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 0, 4, 1, 5),
      wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 2, 6, 3, 7));
    const v128_t vacc3x01c2 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 0, 4, 1, 5),
      wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 2, 6, 3, 7));

    v128_t vacc01x01 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 0, 1, 4, 5),
      wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 2, 3, 6, 7));
    v128_t vacc23x01 = wasm_f32x4_add(
      wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 0, 1, 4, 5),
      wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 2, 3, 6, 7));

    vacc01x01 = wasm_f32x4_pmax(vmin, vacc01x01);
    vacc23x01 = wasm_f32x4_pmax(vmin, vacc23x01);

    vacc01x01 = wasm_f32x4_pmin(vmax, vacc01x01);
    vacc23x01 = wasm_f32x4_pmin(vmax, vacc23x01);

    if XNN_LIKELY(nc >= 2) {
      *((double*) c3) = wasm_f64x2_extract_lane(vacc23x01, 1);
      c3 = (float*) ((uintptr_t) c3 + cn_stride);
      *((double*) c2) = wasm_f64x2_extract_lane(vacc23x01, 0);
      c2 = (float*) ((uintptr_t) c2 + cn_stride);
      *((double*) c1) = wasm_f64x2_extract_lane(vacc01x01, 1);
      c1 = (float*) ((uintptr_t) c1 + cn_stride);
      *((double*) c0) = wasm_f64x2_extract_lane(vacc01x01, 0);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= 2;
    } else {
      assert(nc == 1);
      *c3 = wasm_f32x4_extract_lane(vacc23x01, 2);
      *c2 = wasm_f32x4_extract_lane(vacc23x01, 0);
      *c1 = wasm_f32x4_extract_lane(vacc01x01, 2);
      *c0 = wasm_f32x4_extract_lane(vacc01x01, 0);

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