src/qs8-gavgpool/unipass-neon.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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 3
$assert REQUANTIZATION == "FP32"
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gavgpool.h>
$if ARMV8:
  #include <xnnpack/intrinsics-polyfill.h>


$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if ARMV8 else "neon")
$ISA = "neonv8" if ARMV8 else "neon"
void xnn_qs8_gavgpool_minmax_fp32_ukernel_${ROW_TILE}x__${ISA}_c${CHANNEL_TILE}(
    size_t rows,
    size_t channels,
    const int8_t* input,
    size_t input_stride,
    const int8_t* zero,
    int8_t* output,
    const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows != 0);
  assert(rows <= ${ROW_TILE});
  assert(channels != 0);

  const int8_t* i0 = input;
  $for M in range(1, ROW_TILE):
    const int8_t* i${M} = (const int8_t*) ((uintptr_t) i${M-1} + input_stride);
    $if M % 2 == 1:
      if XNN_UNPREDICTABLE(rows < ${M+1}) {
        i${M} = zero;
      }
    $else:
      if XNN_UNPREDICTABLE(rows <= ${M}) {
        i${M} = zero;
      }

  const int32x4_t vinit_bias = vld1q_dup_s32(&params->${PARAMS_STRUCT}.init_bias);
  const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
  $if ARMV8:
    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
  $else:
    const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
    const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
  $if CHANNEL_TILE > 8:
    const int8x16_t voutput_min = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_min);
    const int8x16_t voutput_max = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_max);
  $else:
    const int8x8_t voutput_min = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_min);
    const int8x8_t voutput_max = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_max);
  for (; channels >= ${CHANNEL_TILE}; channels -= ${CHANNEL_TILE}) {
    $for M in range(2):
      $for C in range(0, CHANNEL_TILE, 8):
        const int8x8_t vi${M}x${ABC[C:C+8]} = vld1_s8(i${M}); i${M} += 8;

    $for C in range(0, CHANNEL_TILE, 8):
      const int8x8_t vi2x${ABC[C:C+8]} = vld1_s8(i2); i2 += 8;
      int16x8_t vacc${ABC[C:C+8]} = vaddl_s8(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});

    $for M in range(2, ROW_TILE):
      $for C in range(0, CHANNEL_TILE, 8):
        $if M + 1 != ROW_TILE:
          const int8x8_t vi${M+1}x${ABC[C:C+8]} = vld1_s8(i${M+1}); i${M+1} += 8;
        vacc${ABC[C:C+8]} = vaddw_s8(vacc${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});

    $for C in range(0, CHANNEL_TILE, 8):
      int32x4_t vacc${ABC[C:C+4]} = vaddw_s16(vinit_bias, vget_low_s16(vacc${ABC[C:C+8]}));
      int32x4_t vacc${ABC[C+4:C+8]} = vaddw_s16(vinit_bias, vget_high_s16(vacc${ABC[C:C+8]}));

    $for C in range(0, CHANNEL_TILE, 4):
      float32x4_t vfpacc${ABC[C:C+4]} = vcvtq_f32_s32(vacc${ABC[C:C+4]});

    $for C in range(0, CHANNEL_TILE, 4):
      vfpacc${ABC[C:C+4]} = vmulq_f32(vfpacc${ABC[C:C+4]}, vscale);

    $if ARMV8:
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vcvtnq_s32_f32(vfpacc${ABC[C:C+4]});
    $else:
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[C:C+4]}, vmagic_bias));

      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = vqsubq_s32(vacc${ABC[C:C+4]}, vmagic_bias_less_output_zero_point);

    #if XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[C:C+4]}), vacc${ABC[C+4:C+8]});
    #else  // !XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[C:C+4]}), vqmovn_s32(vacc${ABC[C+4:C+8]}));
    #endif  // !XNN_ARCH_ARM64

    $if ARMV8:
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = vqaddq_s16(vacc${ABC[C:C+8]}, voutput_zero_point);

    #if XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          int8x16_t vout${ABC[C:C+16]} = vqmovn_high_s16(vqmovn_s16(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]});
        $else:
          int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]});
    #else  // !XNN_ARCH_ARM64
      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          int8x16_t vout${ABC[C:C+16]} = vcombine_s8(vqmovn_s16(vacc${ABC[C:C+8]}), vqmovn_s16(vacc${ABC[C+8:C+16]}));
        $else:
          int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]});
    #endif  // !XNN_ARCH_ARM64

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

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

    $for C in range(0, CHANNEL_TILE, 16):
      $if C + 8 < CHANNEL_TILE:
        vst1q_s8(output, vout${ABC[C:C+16]}); output += 16;
      $else:
        vst1_s8(output, vout${ABC[C:C+8]}); output += 8;
  }
  if XNN_UNLIKELY(channels != 0) {
    ${"do " if CHANNEL_TILE > 8 else ""}{
      $for M in range(3):
        const int8x8_t vi${M}x${ABC[0:8]} = vld1_s8(i${M}); i${M} += 8;
      int16x8_t vacc${ABC[0:8]} = vaddl_s8(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});

      $for M in range(2, ROW_TILE):
        $if M + 1 != ROW_TILE:
          const int8x8_t vi${M+1}x${ABC[0:8]} = vld1_s8(i${M+1}); i${M+1} += 8;
        vacc${ABC[0:8]} = vaddw_s8(vacc${ABC[0:8]}, vi${M}x${ABC[0:8]});

      int32x4_t vacc${ABC[0:4]} = vaddw_s16(vinit_bias, vget_low_s16(vacc${ABC[0:8]}));
      int32x4_t vacc${ABC[4:8]} = vaddw_s16(vinit_bias, vget_high_s16(vacc${ABC[0:8]}));

      float32x4_t vfpacc${ABC[0:4]} = vcvtq_f32_s32(vacc${ABC[0:4]});
      float32x4_t vfpacc${ABC[4:8]} = vcvtq_f32_s32(vacc${ABC[4:8]});

      vfpacc${ABC[0:4]} = vmulq_f32(vfpacc${ABC[0:4]}, vscale);
      vfpacc${ABC[4:8]} = vmulq_f32(vfpacc${ABC[4:8]}, vscale);

      $if ARMV8:
        vacc${ABC[0:4]} = vcvtnq_s32_f32(vfpacc${ABC[0:4]});
        vacc${ABC[4:8]} = vcvtnq_s32_f32(vfpacc${ABC[4:8]});
      $else:
        vacc${ABC[0:4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[0:4]}, vmagic_bias));
        vacc${ABC[4:8]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${ABC[4:8]}, vmagic_bias));

        vacc${ABC[0:4]} = vqsubq_s32(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
        vacc${ABC[4:8]} = vqsubq_s32(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

      #if XNN_ARCH_ARM64
        vacc${ABC[0:8]} = vqmovn_high_s32(vqmovn_s32(vacc${ABC[0:4]}), vacc${ABC[4:8]});
      #else
        vacc${ABC[0:8]} = vcombine_s16(vqmovn_s32(vacc${ABC[0:4]}), vqmovn_s32(vacc${ABC[4:8]}));
      #endif
      $if ARMV8:
        vacc${ABC[0:8]} = vqaddq_s16(vacc${ABC[0:8]}, voutput_zero_point);

      int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
      $if CHANNEL_TILE > 8:
        vout${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 XNN_LIKELY(channels >= 8) {
          vst1_s8(output, vout${ABC[0:8]}); output += 8;
          channels -= 8;
        } else {
          if (channels & 4) {
            vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout${ABC[0:8]}), 0); output += 4;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 4);
          }
          if (channels & 2) {
            vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout${ABC[0:8]}), 0); output += 2;
            vout${ABC[0:8]} = vext_s8(vout${ABC[0:8]}, vout${ABC[0:8]}, 2);
          }
          if (channels & 1) {
            vst1_lane_s8(output, vout${ABC[0:8]}, 0); output += 1;
          }
          channels = 0;
        }
      $else:
        vout${ABC[0:8]} = vmax_s8(vout${ABC[0:8]}, voutput_min);
        vout${ABC[0:8]} = vmin_s8(vout${ABC[0:8]}, voutput_max);

        if (channels & 4) {
          vst1_lane_u32((void*) output, 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 (channels & 2) {
          vst1_lane_u16((void*) output, 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 (channels & 1) {
          vst1_lane_s8(output, vout${ABC[0:8]}, 0);
        }
    }${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
  }
}