xref: /aosp_15_r20/external/XNNPACK/src/s8-ibilinear/gen/wasmsimd-mul32-c8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/s8-ibilinear/wasmsimd-mul32.c.in
//   Generator: tools/xngen
//
// 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 <wasm_simd128.h>

#include <xnnpack/common.h>
#include <xnnpack/ibilinear.h>


void xnn_s8_ibilinear_ukernel__wasmsimd_mul32_c8(
    size_t output_pixels,
    size_t channels,
    const int8_t**restrict input,
    size_t input_offset,
    const int16_t*restrict weights,
    int8_t*restrict output,
    size_t output_increment) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(channels != 0);

  do {
    const int8_t* i0 = (const int8_t*) ((uintptr_t) input[0] + input_offset);
    const int8_t* i1 = (const int8_t*) ((uintptr_t) input[1] + input_offset);
    const int8_t* i2 = (const int8_t*) ((uintptr_t) input[2] + input_offset);
    const int8_t* i3 = (const int8_t*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const v128_t valphah = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights));
    const v128_t valphav = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights + 1));
    weights += 2;

    const v128_t vrounding = wasm_i32x4_const_splat(0x00200000);

    size_t c = channels;
    for (; c >= 8 * sizeof(int8_t); c -= 8 * sizeof(int8_t)) {
      const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
      i0 += 8;
      const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
      i1 += 8;
      const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
      i2 += 8;
      const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);
      i3 += 8;

      const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
      const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
      const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
      const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);

      const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
      const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
      const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
      const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));

      v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
      v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);

      vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
      vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);

      vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
      vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);

      const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);

      const v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);

      *((double*) output) = wasm_f64x2_extract_lane(vo01234567, 0);
      output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
      const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
      const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
      const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);

      const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
      const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
      const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
      const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);

      const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
      const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
      const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
      const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));

      v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
      v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);

      vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
      vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);

      vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
      vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);

      const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);

      v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);

      if (c & (4 * sizeof(int8_t))) {
        *((float*) output) = wasm_f32x4_extract_lane(vo01234567, 0);
        output += 4;
        vo01234567 = wasm_u64x2_shr(vo01234567, 32);
      }
      uint32_t vo0123 = (uint32_t) wasm_i32x4_extract_lane(vo01234567, 0);
      if (c & (2 * sizeof(int8_t))) {
        *((uint16_t*) output) = (uint16_t) vo0123;
        output += 2;
        vo0123 >>= 16;
      }
      if (c & (1 * sizeof(int8_t))) {
        *output++ = (uint8_t) vo0123;
      }
    }

    output = (int8_t*) ((uintptr_t) output + output_increment);
  } while (--output_pixels != 0);
}