src/x24-transposec/4x4-aarch64-tbl.c

// Copyright 2022 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 <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/transpose.h>

static const uint8_t pos0[16] = {0, 1, 2, 16, 17, 18, 32, 33, 34, 48, 49, 50, 0, 0, 0, 0};
static const uint8_t pos1[16] = {3, 4, 5, 19, 20, 21, 35, 36, 37, 51, 52, 53, 0, 0, 0, 0};
static const uint8_t pos2[16] = {6, 7, 8, 22, 23, 24, 38, 39, 40, 54, 55, 56, 0, 0, 0, 0};
static const uint8_t pos3[16] = {9, 10, 11, 25, 26, 27, 41, 42, 43, 57, 58, 59, 0, 0, 0, 0};

void xnn_x24_transposec_ukernel__4x4_aarch64_neon_tbl(
    const void* input,
    void* output,
    size_t input_stride,
    size_t output_stride,
    size_t block_width,
    size_t block_height)
{
  assert(output_stride >= block_height * 3);
  assert(input_stride >= block_width * 3);

  const size_t tile_height = 4;
  const size_t tile_width = 4;
  const size_t tile_wbytes = tile_width * 3;
  const size_t tile_wbytes_minus_8 = tile_wbytes - 8;
  const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
  const size_t output_reset = tile_width * output_stride - block_height * 3;
  const size_t tile_stride = tile_height * input_stride;

  const uint8_t* i0 = (const uint8_t*) input;
  const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
  const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
  const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);

  uint8_t* o0 = (uint8_t*) output;
  uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
  uint8_t* o2 = (uint8_t*) ((uintptr_t) o1 + output_stride);
  uint8_t* o3 = (uint8_t*) ((uintptr_t) o2 + output_stride);

  const uint8x16_t vperm0 = vld1q_u8(pos0);
  const uint8x16_t vperm1 = vld1q_u8(pos1);
  const uint8x16_t vperm2 = vld1q_u8(pos2);
  const uint8x16_t vperm3 = vld1q_u8(pos3);
  do {
    if XNN_UNPREDICTABLE(block_width < 2) {
      o1 = o0;
    }
    if XNN_UNPREDICTABLE(block_width <= 2) {
      o2 = o0;
    }
    if XNN_UNPREDICTABLE(block_width < 4) {
      o3 = o0;
    }
    size_t bh = block_height;
    for (; bh >= 4; bh -= 4) {
      uint8x16x4_t v;
      v.val[0] = vld1q_u8(i0); i0 = (const uint8_t*) ((uintptr_t) i0 + tile_stride);
      v.val[1] = vld1q_u8(i1); i1 = (const uint8_t*) ((uintptr_t) i1 + tile_stride);
      v.val[2] = vld1q_u8(i2); i2 = (const uint8_t*) ((uintptr_t) i2 + tile_stride);
      v.val[3] = vld1q_u8(i3); i3 = (const uint8_t*) ((uintptr_t) i3 + tile_stride);

      const uint8x16_t vres0 = vqtbl4q_u8(v, vperm0);
      const uint8x16_t vres1 = vqtbl4q_u8(v, vperm1);
      const uint8x16_t vres2 = vqtbl4q_u8(v, vperm2);
      const uint8x16_t vres3 = vqtbl4q_u8(v, vperm3);

      vst1_u8(o3, vget_low_u8(vres3)); o3 += 8;
      vst1_u8(o2, vget_low_u8(vres2)); o2 += 8;
      vst1_u8(o1, vget_low_u8(vres1)); o1 += 8;
      vst1_u8(o0, vget_low_u8(vres0)); o0 += 8;
      vst1q_lane_u32((void*) o3, vreinterpretq_u32_u8(vres3), 2); o3 = (uint8_t*) ((uintptr_t) o3 + tile_wbytes_minus_8);
      vst1q_lane_u32((void*) o2, vreinterpretq_u32_u8(vres2), 2); o2 = (uint8_t*) ((uintptr_t) o2 + tile_wbytes_minus_8);
      vst1q_lane_u32((void*) o1, vreinterpretq_u32_u8(vres1), 2); o1 = (uint8_t*) ((uintptr_t) o1 + tile_wbytes_minus_8);
      vst1q_lane_u32((void*) o0, vreinterpretq_u32_u8(vres0), 2); o0 = (uint8_t*) ((uintptr_t) o0 + tile_wbytes_minus_8);
    }

    if (bh != 0) {
      if XNN_UNPREDICTABLE(bh <= 2) {
        i2 = i0;
      }
      if XNN_UNPREDICTABLE(bh < 2) {
        i1 = i0;
      }
      uint8x16x3_t v;
      v.val[0] = vld1q_u8(i0);
      v.val[1] = vld1q_u8(i1);
      v.val[2] = vld1q_u8(i2);

      uint8x16_t vres0 = vqtbl3q_u8(v, vperm0);
      uint8x16_t vres1 = vqtbl3q_u8(v, vperm1);
      uint8x16_t vres2 = vqtbl3q_u8(v, vperm2);
      uint8x16_t vres3 = vqtbl3q_u8(v, vperm3);

      uint8x8_t vres0_lo = vget_low_u8(vres0);
      uint8x8_t vres1_lo = vget_low_u8(vres1);
      uint8x8_t vres2_lo = vget_low_u8(vres2);
      uint8x8_t vres3_lo = vget_low_u8(vres3);

      if (bh & 2) {
        vst1_lane_u32((void*) o3, vreinterpret_u32_u8(vres3_lo), 0); o3 += 4;
        vst1_lane_u32((void*) o2, vreinterpret_u32_u8(vres2_lo), 0); o2 += 4;
        vst1_lane_u32((void*) o1, vreinterpret_u32_u8(vres1_lo), 0); o1 += 4;
        vst1_lane_u32((void*) o0, vreinterpret_u32_u8(vres0_lo), 0); o0 += 4;
        vst1_lane_u16((void*) o3, vreinterpret_u16_u8(vres3_lo), 2); o3 += 2;
        vst1_lane_u16((void*) o2, vreinterpret_u16_u8(vres2_lo), 2); o2 += 2;
        vst1_lane_u16((void*) o1, vreinterpret_u16_u8(vres1_lo), 2); o1 += 2;
        vst1_lane_u16((void*) o0, vreinterpret_u16_u8(vres0_lo), 2); o0 += 2;
        vres0_lo = vget_low_u8(vextq_u8(vres0, vres0, 6));
        vres1_lo = vget_low_u8(vextq_u8(vres1, vres1, 6));
        vres2_lo = vget_low_u8(vextq_u8(vres2, vres2, 6));
        vres3_lo = vget_low_u8(vextq_u8(vres3, vres3, 6));
      }
      if (bh & 1) {
        vst1_lane_u16((void*) o3, vreinterpret_u16_u8(vres3_lo), 0); o3 += 2;
        vst1_lane_u16((void*) o2, vreinterpret_u16_u8(vres2_lo), 0); o2 += 2;
        vst1_lane_u16((void*) o1, vreinterpret_u16_u8(vres1_lo), 0); o1 += 2;
        vst1_lane_u16((void*) o0, vreinterpret_u16_u8(vres0_lo), 0); o0 += 2;
        vst1_lane_u8(o3, vres3_lo, 2); o3 += 1;
        vst1_lane_u8(o2, vres2_lo, 2); o2 += 1;
        vst1_lane_u8(o1, vres1_lo, 2); o1 += 1;
        vst1_lane_u8(o0, vres0_lo, 2); o0 += 1;
      }
    }
    i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
    i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
    i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
    i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
    o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
    o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
    o2 = (uint8_t*) ((uintptr_t) o2 + output_reset);
    o3 = (uint8_t*) ((uintptr_t) o3 + output_reset);
    block_width = doz(block_width, tile_width);
  } while (block_width != 0);
}