src/s16-window/neon.c.in

// 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.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include <arm_neon.h>

#include <xnnpack/math.h>
#include <xnnpack/window.h>

$SHIFT_VARIANT = "_shift%s" % SHIFT if SHIFT else ""

void xnn_s16_window${SHIFT_VARIANT}_ukernel__neon_x${BATCH_TILE}(
    size_t rows,
    size_t batch_size,
    const int16_t* input,
    const int16_t* weights,
    int16_t* output,
    uint32_t shift)
{
  assert(rows != 0);
  assert(batch_size != 0);
  assert(input != NULL);
  assert(weights != NULL);
  assert(output != NULL);
  $if SHIFT != 0:
    assert(shift == ${SHIFT});
  $else:
    assert(shift < 32);

  $if SHIFT == 0:
    const int32x4_t vshift = vdupq_n_s32(-(int32_t)shift);  // negative to shift right.

  do {
    const int16_t* w = weights;
    size_t n = batch_size * sizeof(int16_t);
    $if BATCH_TILE > 8:
      for (; n >= ${BATCH_TILE} * sizeof(int16_t); n -= ${BATCH_TILE} * sizeof(int16_t)) {
        $for N in range(SIMD_TILE):
          const int16x8_t vi${N} = vld1q_s16(input); input += 8;

        $for N in range(SIMD_TILE):
          const int16x8_t vw${N} = vld1q_s16(w); w += 8;

        $if SHIFT == 15:
          $for N in range(SIMD_TILE):
            const int16x8_t vout${N} = vqdmulhq_s16(vi${N}, vw${N});
        $else:
          $for N in range(SIMD_TILE):
            int32x4_t vacc${N}_lo = vmull_s16(vget_low_s16(vi${N}), vget_low_s16(vw${N}));
            int32x4_t vacc${N}_hi = vmull_s16(vget_high_s16(vi${N}), vget_high_s16(vw${N}));

          $if SHIFT != 0:
            $for N in range(SIMD_TILE):
              const int16x4_t vshift${N}_lo = vqshrn_n_s32(vacc${N}_lo, ${SHIFT});
              const int16x4_t vshift${N}_hi = vqshrn_n_s32(vacc${N}_hi, ${SHIFT});

            $for N in range(SIMD_TILE):
              const int16x8_t vout${N} = vcombine_s16(vshift${N}_lo, vshift${N}_hi);
          $else:
            $for N in range(SIMD_TILE):
              vacc${N}_lo = vshlq_s32(vacc${N}_lo, vshift);
              vacc${N}_hi = vshlq_s32(vacc${N}_hi, vshift);

            $for N in range(SIMD_TILE):
              const int16x8_t vout${N} = vcombine_s16(vqmovn_s32(vacc${N}_lo), vqmovn_s32(vacc${N}_hi));

        $for N in range(SIMD_TILE):
          vst1q_s16(output, vout${N}); output += 8;
      }

    // Remainder of full vectors
    for (; n >= 8 * sizeof(int16_t); n -= 8 * sizeof(int16_t)) {
      const int16x8_t vi = vld1q_s16(input); input += 8;
      const int16x8_t vw = vld1q_s16(w); w += 8;
      $if SHIFT == 15:
        const int16x8_t vout = vqdmulhq_s16(vi, vw);
      $else:
        int32x4_t vacc_lo = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
        int32x4_t vacc_hi = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
        $if SHIFT != 0:
          const int16x4_t vshift_lo = vqshrn_n_s32(vacc_lo, ${SHIFT});
          const int16x4_t vshift_hi = vqshrn_n_s32(vacc_hi, ${SHIFT});
          const int16x8_t vout = vcombine_s16(vshift_lo, vshift_hi);
        $else:
          vacc_lo = vshlq_s32(vacc_lo, vshift);
          vacc_hi = vshlq_s32(vacc_hi, vshift);
          const int16x8_t vout = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
      vst1q_s16(output, vout); output += 8;
    }

    assert(n % 2 == 0);
    // Remainder of 1 to 7 batch_size
    if XNN_UNLIKELY(n != 0) {
      const int16x8_t vi = vld1q_s16(input); input = (const int16_t*) ((uintptr_t) input + n);
      const int16x8_t vw = vld1q_s16(w);
      $if SHIFT == 15:
        int16x4_t vout = vqdmulh_s16(vget_low_s16(vi), vget_low_s16(vw));
      $else:
        int32x4_t vacc = vmull_s16(vget_low_s16(vi), vget_low_s16(vw));
        $if SHIFT != 0:
          int16x4_t vout = vqshrn_n_s32(vacc, ${SHIFT});
        $else:
          vacc = vshlq_s32(vacc, vshift);
          int16x4_t vout = vqmovn_s32(vacc);
      if (n & (4 * sizeof(int16_t))) {
        vst1_s16(output, vout); output += 4;
        $if SHIFT == 15:
          vout = vqdmulh_s16(vget_high_s16(vi), vget_high_s16(vw));
        $else:
          vacc = vmull_s16(vget_high_s16(vi), vget_high_s16(vw));
          $if SHIFT != 0:
            vout = vqshrn_n_s32(vacc, ${SHIFT});
          $else:
            vacc = vshlq_s32(vacc, vshift);
            vout = vqmovn_s32(vacc);
      }
      if (n & (2 * sizeof(int16_t))) {
        vst1_lane_u32((void*) output, vreinterpret_u32_s16(vout), 0); output += 2;
        vout = vext_s16(vout, vout, 2);
      }
      if (n & (1 * sizeof(int16_t))) {
        vst1_lane_s16(output, vout, 0); output += 1;
      }
    }

  } while (--rows != 0);
}