src/qs8-vaddc/neon-ld64.c.in

// 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.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/vadd.h>

#include <stdio.h>
#include <inttypes.h>


void xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x${BATCH_TILE}(
    size_t n,
    const int8_t* input_x,
    const int8_t* input_y,
    int8_t* output,
    const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  const int8x8_t vx_zero_point = vld1_dup_s8(&params->neon.x_zero_point);
  const int32x4_t vx_multiplier = vld1q_dup_s32(&params->neon.x_multiplier);
  const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
  const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
  const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
  const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);

  const int32_t vey = (int32_t) *input_y - (int32_t) params->neon.y_zero_point;
  const int32_t vy_multiplier = params->neon.y_multiplier;
  const int32x4_t vy_bias = vdupq_n_s32(vey * vy_multiplier);

  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
    $for N in range(0, BATCH_TILE, 8):
      const int8x8_t vx${ABC[N:N+8]} = vld1_s8(input_x); input_x += 8;

    $for N in range(0, BATCH_TILE, 8):
      const int16x8_t vex${ABC[N:N+8]} = vsubl_s8(vx${ABC[N:N+8]}, vx_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      int32x4_t vacc${ABC[N:N+4]} = vmlaq_s32(vy_bias, vmovl_s16(vget_low_s16(vex${ABC[N:N+8]})), vx_multiplier);
      int32x4_t vacc${ABC[N+4:N+8]} = vmlaq_s32(vy_bias, vmovl_s16(vget_high_s16(vex${ABC[N:N+8]})), vx_multiplier);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = vsraq_n_s32(vacc${ABC[N:N+4]}, vbicq_s32(vacc${ABC[N:N+4]}, vzero_shift_mask), 31);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = vrshlq_s32(vacc${ABC[N:N+4]}, vright_shift);

    $for N in range(0, BATCH_TILE, 8):
      const int16x8_t vacc${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[N:N+4]}), vqmovn_s32(vacc${ABC[N+4:N+8]})), voutput_zero_point);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        int8x16_t vout${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${ABC[N:N+8]}), vqmovn_s16(vacc${ABC[N+8:N+16]}));
      $else:
        int8x8_t vout${ABC[N:N+8]} = vqmovn_s16(vacc${ABC[N:N+8]});

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = vmaxq_s8(vout${ABC[N:N+16]}, voutput_min);
      $else:
        vout${ABC[N:N+8]} = vmax_s8(vout${ABC[N:N+8]}, vget_low_s8(voutput_min));

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = vminq_s8(vout${ABC[N:N+16]}, voutput_max);
      $else:
        vout${ABC[N:N+8]} = vmin_s8(vout${ABC[N:N+8]}, vget_low_s8(voutput_max));

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vst1q_s8(output, vout${ABC[N:N+16]}); output += 16;
      $else:
        vst1_s8(output, vout${ABC[N:N+8]}); output += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    ${"do " if BATCH_TILE > 8 else ""}{
      $if BATCH_TILE > 8:
        const int8x8_t vx${ABC[0:8]} = vld1_s8(input_x); input_x += 8;
      $else:
        const int8x8_t vx${ABC[0:8]} = vld1_s8(input_x);

      const int16x8_t vex${ABC[0:8]} = vsubl_s8(vx${ABC[0:8]}, vx_zero_point);

      int32x4_t vacc${ABC[0:4]} = vmlaq_s32(vy_bias, vmovl_s16(vget_low_s16(vex${ABC[0:8]})), vx_multiplier);
      int32x4_t vacc${ABC[4:8]} = vmlaq_s32(vy_bias, vmovl_s16(vget_high_s16(vex${ABC[0:8]})), vx_multiplier);

      vacc${ABC[0:4]} = vsraq_n_s32(vacc${ABC[0:4]}, vbicq_s32(vacc${ABC[0:4]}, vzero_shift_mask), 31);
      vacc${ABC[4:8]} = vsraq_n_s32(vacc${ABC[4:8]}, vbicq_s32(vacc${ABC[4:8]}, vzero_shift_mask), 31);

      vacc${ABC[0:4]} = vrshlq_s32(vacc${ABC[0:4]}, vright_shift);
      vacc${ABC[4:8]} = vrshlq_s32(vacc${ABC[4:8]}, vright_shift);

      const int16x8_t vacc${ABC[0:8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]})), voutput_zero_point);

      int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
      vout${ABC[0:8]} = vmax_s8(vout${ABC[0:8]}, vget_low_s8(voutput_min));
      vout${ABC[0:8]} = vmin_s8(vout${ABC[0:8]}, vget_low_s8(voutput_max));

      $if BATCH_TILE > 8:
        if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
          vst1_s8(output, vout${ABC[0:8]}); output += 8;
          n -= 8 * sizeof(int8_t);
        } else {
          if (n & (4 * sizeof(int8_t))) {
            vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_s8(vout${ABC[0:8]}), 0); output += 4;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
          }
          if (n & (2 * sizeof(int8_t))) {
            vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_s8(vout${ABC[0:8]}), 0); output += 2;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
          }
          if (n & (1 * sizeof(int8_t))) {
            vst1_lane_s8(output, vout${ABC[0:8]}, 0);
          }
          n = 0;
        }
      $else:
        if (n & (4 * sizeof(int8_t))) {
          vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_s8(vout${ABC[0:8]}), 0); output += 4;
          vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
        }
        if (n & (2 * sizeof(int8_t))) {
          vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_s8(vout${ABC[0:8]}), 0); output += 2;
          vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
        }
        if (n & (1 * sizeof(int8_t))) {
          vst1_lane_s8(output, vout${ABC[0:8]}, 0);
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}