src/qs8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-mul16-ld128.c

// Auto-generated file. Do not edit!
//   Template: src/qs8-igemm/MRx4c8-wasmsimd-mul16.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/gemm.h>
#include <xnnpack/math.h>


void xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const int8_t** restrict a,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const int8_t* zero,
    const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
{
  assert(mr != 0);
  assert(mr <= 3);
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (3 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8);
  int8_t* c0 = c;
  int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    c1 = c0;
  }
  int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    c2 = c1;
  }

  do {
    v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
    v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
    v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
    v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
    v128_t vacc1x0 = vacc0x0;
    v128_t vacc1x1 = vacc0x1;
    v128_t vacc1x2 = vacc0x2;
    v128_t vacc1x3 = vacc0x3;
    v128_t vacc2x0 = vacc0x0;
    v128_t vacc2x1 = vacc0x1;
    v128_t vacc2x2 = vacc0x2;
    v128_t vacc2x3 = vacc0x3;
    w = (const void*) ((const int32_t*) w + 4);

    size_t p = ks;
    do {
      const int8_t* restrict a0 = a[0];
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
      }
      const int8_t* restrict a1 = a[1];
      if XNN_UNPREDICTABLE(a1 != zero) {
        a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
      }
      const int8_t* restrict a2 = a[2];
      if XNN_UNPREDICTABLE(a2 != zero) {
        a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
      }
      a += 3;

      size_t k = 0;
      while (k < kc) {
        const v128_t vxa0 = wasm_i16x8_load8x8(a0);
        a0 += 8;
        const v128_t vxa1 = wasm_i16x8_load8x8(a1);
        a1 += 8;
        const v128_t vxa2 = wasm_i16x8_load8x8(a2);
        a2 += 8;

        const v128_t vb01 = wasm_v128_load(w);
        const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
        const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);

        const v128_t vprod0x0 = wasm_i16x8_mul(vxb0, vxa0);
        vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_low_i16x8(vprod0x0));
        const v128_t vprod1x0 = wasm_i16x8_mul(vxb0, vxa1);
        vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_extend_low_i16x8(vprod1x0));
        const v128_t vprod2x0 = wasm_i16x8_mul(vxb0, vxa2);
        vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_extend_low_i16x8(vprod2x0));

        const v128_t vprod0x1 = wasm_i16x8_mul(vxb1, vxa0);
        vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_low_i16x8(vprod0x1));
        vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_high_i16x8(vprod0x0));
        const v128_t vprod1x1 = wasm_i16x8_mul(vxb1, vxa1);
        vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_extend_low_i16x8(vprod1x1));
        vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_extend_high_i16x8(vprod1x0));
        const v128_t vprod2x1 = wasm_i16x8_mul(vxb1, vxa2);
        vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_extend_low_i16x8(vprod2x1));
        vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_extend_high_i16x8(vprod2x0));

        vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_high_i16x8(vprod0x1));
        vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_extend_high_i16x8(vprod1x1));
        vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_extend_high_i16x8(vprod2x1));
        const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
        const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
        const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);

        const v128_t vprod0x2 = wasm_i16x8_mul(vxb2, vxa0);
        vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_low_i16x8(vprod0x2));
        const v128_t vprod1x2 = wasm_i16x8_mul(vxb2, vxa1);
        vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_extend_low_i16x8(vprod1x2));
        const v128_t vprod2x2 = wasm_i16x8_mul(vxb2, vxa2);
        vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_extend_low_i16x8(vprod2x2));

        const v128_t vprod0x3 = wasm_i16x8_mul(vxb3, vxa0);
        vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_low_i16x8(vprod0x3));
        vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_high_i16x8(vprod0x2));
        const v128_t vprod1x3 = wasm_i16x8_mul(vxb3, vxa1);
        vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_extend_low_i16x8(vprod1x3));
        vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_extend_high_i16x8(vprod1x2));
        const v128_t vprod2x3 = wasm_i16x8_mul(vxb3, vxa2);
        vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_extend_low_i16x8(vprod2x3));
        vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_extend_high_i16x8(vprod2x2));

        vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_high_i16x8(vprod0x3));
        vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_extend_high_i16x8(vprod1x3));
        vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_extend_high_i16x8(vprod2x3));

        w = (const void*) ((const int8_t*) w + 32);
        k += 8 * sizeof(int8_t);
      }
      p -= 3 * sizeof(void*);
    } while (p != 0);

    const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
    const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
    const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
    const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
    const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
    const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));

    v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
    v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
    v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));

    vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
    vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
    vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);

    const v128_t vscale = wasm_v128_load(params->fp32_wasmsimd.scale);
    vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
    vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
    vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);

    const v128_t voutput_min_less_zero_point = wasm_v128_load(params->fp32_wasmsimd.output_min_less_zero_point);
    vacc0x0123 = wasm_f32x4_pmax(voutput_min_less_zero_point, vacc0x0123);
    vacc1x0123 = wasm_f32x4_pmax(voutput_min_less_zero_point, vacc1x0123);
    vacc2x0123 = wasm_f32x4_pmax(voutput_min_less_zero_point, vacc2x0123);

    const v128_t voutput_max_less_zero_point = wasm_v128_load(params->fp32_wasmsimd.output_max_less_zero_point);
    vacc0x0123 = wasm_f32x4_pmin(voutput_max_less_zero_point, vacc0x0123);
    vacc1x0123 = wasm_f32x4_pmin(voutput_max_less_zero_point, vacc1x0123);
    vacc2x0123 = wasm_f32x4_pmin(voutput_max_less_zero_point, vacc2x0123);

    const v128_t vmagic_bias = wasm_v128_load(params->fp32_wasmsimd.magic_bias);
    vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
    vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
    vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);

    const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
    vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
    vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
    vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);

    v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
    v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);

    v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);

    if (nc >= 4) {
      *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
      *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
      *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);

      c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
      c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

      a = (const int8_t**restrict) ((uintptr_t) a - ks);

      nc -= 4;
    } else {
      uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
      uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
      uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
      if (nc & 2) {
        *((uint16_t*) c2) = (uint16_t) vout2;
        vout2 >>= 16;
        c2 += 2;
        *((uint16_t*) c1) = (uint16_t) vout1;
        vout1 >>= 16;
        c1 += 2;
        *((uint16_t*) c0) = (uint16_t) vout0;
        vout0 >>= 16;
        c0 += 2;
      }
      if (nc & 1) {
        *c2 = (int8_t) vout2;
        *c1 = (int8_t) vout1;
        *c0 = (int8_t) vout0;
      }

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