// 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.

$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert NR % 8 == 0
$assert 8 <= NR <= 16
$assert REQUANTIZATION in ["FP32", "RNDNU"]
$assert not CHANNELWISE or REQUANTIZATION == "FP32"
$assert DUP in ["DUP", "LD1R", "LD2R", "LD4R"]
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gemm.h>
$if REQUANTIZATION == "FP32" and ARMV8:
  #include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>

$DATATYPE = "qc8" if CHANNELWISE else "qs8"
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if REQUANTIZATION == "FP32" and ARMV8 else "neon")
$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
$ISA = "neonv8" if ARMV8 else "neon"
void xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c2__${ISA}_${"mlal" if MLA else "mull"}_${DUP.lower()}(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 2 * sizeof(int8_t));
  const int8_t* a0 = a;
  int8_t* c0 = c;
  $for M in range(1, MR):
    const int8_t* a${M} = (const int8_t*) ((uintptr_t) a${M-1} + a_stride);
    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  do {
    $for N in range(0, NR, 4):
      int32x4_t vacc0x${ABC[N:N+4]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
    $for M in range(1, MR):
      $for N in range(0, NR, 4):
        int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};

    size_t k = kc;

    $if MLA:
      while (k >= 16 * sizeof(int8_t)) {
        $for M in range(MR):
          $if DUP == "LD4R":
            const int16x4x4_t va${M}x0 = vld4_dup_s16((const void*)a${M}); a${M} += 8;
            const int16x4x4_t va${M}x1 = vld4_dup_s16((const void*)a${M}); a${M} += 8;
          $elif DUP == "LD2R":
            const int16x4x2_t va${M}0x0 = vld2_dup_s16((const void*)a${M});
            const int16x4x2_t va${M}1x0 = vld2_dup_s16((const void*)(a${M} + 4)); a${M} += 8;
            const int16x4x2_t va${M}0x1 = vld2_dup_s16((const void*)a${M});
            const int16x4x2_t va${M}1x1 = vld2_dup_s16((const void*)(a${M} + 4)); a${M} += 8;
          $elif DUP == "LD1R":
            const int16x4_t va${M}0x0 = vld1_dup_s16((const void*)a${M});
            const int16x4_t va${M}1x0 = vld1_dup_s16((const void*)(a${M} + 2));
            const int16x4_t va${M}2x0 = vld1_dup_s16((const void*)(a${M} + 4));
            const int16x4_t va${M}3x0 = vld1_dup_s16((const void*)(a${M} + 6)); a${M} += 8;
            const int16x4_t va${M}0x1 = vld1_dup_s16((const void*)a${M});
            const int16x4_t va${M}1x1 = vld1_dup_s16((const void*)(a${M} + 2));
            const int16x4_t va${M}2x1 = vld1_dup_s16((const void*)(a${M} + 4));
            const int16x4_t va${M}3x1 = vld1_dup_s16((const void*)(a${M} + 6)); a${M} += 8;
          $else:
            const int8x8_t va${M}x0 = vld1_s8(a${M}); a${M} += 8;
            const int8x8_t va${M}x1 = vld1_s8(a${M}); a${M} += 8;
        $for K in range(4):
          $for N in range(0, NR, 4):
            const int8x8_t vb${ABC[N:N+4]}c${K}x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

        $for K in range(4):
          $for M in range(MR):
            $if DUP == "LD4R":
              const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s16(va${M}x0.val[${K}]);
              const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s16(va${M}x1.val[${K}]);
            $elif DUP == "LD2R":
              const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s16(va${M}${int(K/2)}x0.val[${K%2}]);
              const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s16(va${M}${int(K/2)}x1.val[${K%2}]);
            $elif DUP == "LD1R":
              const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s16(va${M}${K}x0);
              const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s16(va${M}${K}x1);
            $else:
              const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}x0), ${K}));
              const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}x1), ${K}));

          $for N in range(0, NR, 4):
            $for M in range(MR):
              int16x8_t vprod${M}x${ABC[N:N+4]}c${K} = vmull_s8(vb${ABC[N:N+4]}c${K}x0, va${M}c${K}x0);
            const int8x8_t vb${ABC[N:N+4]}c${K}x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
            $for M in range(MR):
              vprod${M}x${ABC[N:N+4]}c${K} = vmlal_s8(vprod${M}x${ABC[N:N+4]}c${K}, vb${ABC[N:N+4]}c${K}x1, va${M}c${K}x1);
            $for M in range(MR):
              vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c${K});

        k -= 16 * sizeof(int8_t);
      }

    ${"if" if MLA else "while"} (k >= 8 * sizeof(int8_t)) {
      $for M in range(MR):
        $if DUP == "LD4R":
          const int16x4x4_t va${M} = vld4_dup_s16((const void*)a${M}); a${M} += 8;
        $elif DUP == "LD2R":
          const int16x4x2_t va${M}0 = vld2_dup_s16((const void*)a${M});
          const int16x4x2_t va${M}1 = vld2_dup_s16((const void*)(a${M} + 4)); a${M} += 8;
        $elif DUP == "LD1R":
          const int16x4_t va${M}0 = vld1_dup_s16((const void*)a${M});
          const int16x4_t va${M}1 = vld1_dup_s16((const void*)(a${M} + 2));
          const int16x4_t va${M}2 = vld1_dup_s16((const void*)(a${M} + 4));
          const int16x4_t va${M}3 = vld1_dup_s16((const void*)(a${M} + 6)); a${M} += 8;
        $else:
          const int8x8_t va${M} = vld1_s8(a${M}); a${M} += 8;

      $for K in range(4):
        $for N in range(0, NR, 4):
          const int8x8_t vb${ABC[N:N+4]}c${K} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

      $for K in range(4):
        $for M in range(MR):
          $if DUP == "LD4R":
            const int8x8_t va${M}c${K} = vreinterpret_s8_s16(va${M}.val[${K}]);
          $elif DUP == "LD2R":
            const int8x8_t va${M}c${K} = vreinterpret_s8_s16(va${M}${int(K/2)}.val[${K%2}]);
          $elif DUP == "LD1R":
            const int8x8_t va${M}c${K} = vreinterpret_s8_s16(va${M}${K});
          $else:
            const int8x8_t va${M}c${K} = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), ${K}));

        $for N in range(0, NR, 4):
          $for M in range(MR):
            const int16x8_t vprod${M}x${ABC[N:N+4]}c${K} = vmull_s8(vb${ABC[N:N+4]}c${K}, va${M}c${K});
          $for M in range(MR):
            vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c${K});

      k -= 8 * sizeof(int8_t);
    }

    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const int8x8_t va${M} = vld1_s8(a${M}); a${M} = (const int8_t*) ((uintptr_t) a${M} + k);

      $for N in range(0, NR, 4):
        const int8x8_t vb${ABC[N:N+4]}c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

      $for M in range(MR):
        const int8x8_t va${M}c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 0));
        $for N in range(0, NR, 4):
          const int16x8_t vprod${M}x${ABC[N:N+4]}c0 = vmull_s8(vb${ABC[N:N+4]}c0, va${M}c0);
          vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c0);

      if (k > 2 * sizeof(int8_t)) {
        $for N in range(0, NR, 4):
          const int8x8_t vb${ABC[N:N+4]}c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

        $for M in range(MR):
          const int8x8_t va${M}c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 1));
          $for N in range(0, NR, 4):
            const int16x8_t vprod${M}x${ABC[N:N+4]}c1 = vmull_s8(vb${ABC[N:N+4]}c1, va${M}c1);
            vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c1);

        if (k > 4 * sizeof(int8_t)) {
          $for N in range(0, NR, 4):
            const int8x8_t vb${ABC[N:N+4]}c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));

          $for M in range(MR):
            const int8x8_t va${M}c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 2));
            $for N in range(0, NR, 4):
              const int16x8_t vprod${M}x${ABC[N:N+4]}c2 = vmull_s8(vb${ABC[N:N+4]}c2, va${M}c2);
              vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c2);
        }
      }
    }

    $if REQUANTIZATION == "RNDNU":
      const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
      const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
      const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          vacc${M}x${ABC[N:N+4]} = vqshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);

      $for M in range(MR):
        $for N in range(0, NR, 4):
          vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
    $elif REQUANTIZATION == "FP32":
      $for M in range(MR):
        $for N in range(0, NR, 4):
          float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});

      $if CHANNELWISE:
        $for N in range(0, NR, 4):
          const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
          $for M in range(MR):
            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
      $else:
        const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
        $for M in range(MR):
          $for N in range(0, NR, 4):
            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale);

      $if ARMV8:
        $for M in range(MR):
          $for N in range(0, NR, 4):
            vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]});
      $else:
        const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
        $for M in range(MR):
          $for N in range(0, NR, 4):
            vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${M}x${ABC[N:N+4]}, vmagic_bias));

        const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
        $for M in range(MR):
          $for N in range(0, NR, 4):
            vacc${M}x${ABC[N:N+4]} = vqsubq_s32(vacc${M}x${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);

    $if REQUANTIZATION != "FP32" or ARMV8:
      const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
#if XNN_ARCH_ARM64
    $for M in range(MR):
      $for N in range(0, NR, 8):
        int16x8_t vacc${M}x${ABC[N:N+8]} = vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]});

    $if REQUANTIZATION != "FP32" or ARMV8:
      $for M in range(MR):
        $for N in range(0, NR, 8):
          vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);

    $for M in range(MR):
      $for N in range(0, NR, 16):
        $if N + 8 < NR:
          int8x16_t vout${M}x${ABC[N:N+16]} = vqmovn_high_s16(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]});
        $elif M % 2 == 1:
          int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovn_high_s16(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]});
        $elif M + 1 == MR:
          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
#else
    $for M in range(MR):
      $for N in range(0, NR, 8):
        int16x8_t vacc${M}x${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]}));

    $if REQUANTIZATION != "FP32" or ARMV8:
      $for M in range(MR):
        $for N in range(0, NR, 8):
          vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);

    $for M in range(MR):
      $for N in range(0, NR, 16):
        $if N + 8 < NR:
          int8x16_t vout${M}x${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N+8:N+16]}));
        $elif M % 2 == 1:
          int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_s8(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N:N+8]}));
        $elif M + 1 == MR:
          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
#endif

    $if NR == 8 and MR == 1:
      const int8x8_t voutput_min = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_min);
    $else:
      const int8x16_t voutput_min = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_min);
    $for M in range(MR):
      $for N in range(0, NR, 16):
        $if N + 8 < NR:
          vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min);
        $elif M % 2 == 1:
          vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min);
        $elif M + 1 == MR:
          $if NR == 8 and MR == 1:
            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min);
          $else:
            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min));

    $if NR == 8 and MR == 1:
      const int8x8_t voutput_max = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_max);
    $else:
      const int8x16_t voutput_max = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_max);
    $for M in range(MR):
      $for N in range(0, NR, 16):
        $if N + 8 < NR:
          vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max);
        $elif M % 2 == 1:
          vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max);
        $elif M + 1 == MR:
          $if NR == 8 and MR == 1:
            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max);
          $else:
            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max));

    if (nc >= ${NR}) {
      $for M in range(MR):
        $for N in range(0, NR, 16):
          $if N + 8 < NR:
            vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
          $elif M % 2 == 1:
            vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
            vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
          $elif M + 1 == MR:
            vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});

      $for M in range(MR):
        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};
    } else {
      // Final case where not all of the ${NR} columns fit in the destination.
      $if NR == 16:
        $for M in range(MR):
          $if M % 2 == 1:
            int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF));
          $elif M + 1 == MR:
            int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF);
        if (nc & 8) {
          $for M in range(MR):
            $if M % 2 == 1:
              vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
              vst1_s8(c${M}, vget_high_s8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
            $elif M + 1 == MR:
              vst1_s8(c${M}, vout${M}x01234567); c${M} += 8;
          $for M in range(MR):
            $if M % 2 == 1:
              vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF));
            $elif M + 1 == MR:
              vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF);
        }
      if (nc & 4) {
        $for M in range(MR):
          $if M % 2 == 1:
            vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
            vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
          $elif M + 1 == MR:
            vst1_lane_u32((void*) c${M}, vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4;
        $for M in range(MR):
          $if M % 2 == 1:
            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
          $elif M + 1 == MR:
            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4);
      }
      if (nc & 2) {
        $for M in range(MR):
          $if M % 2 == 1:
            vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
            vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
          $elif M + 1 == MR:
            vst1_lane_u16((void*) c${M}, vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2;
        $for M in range(MR):
          $if M % 2 == 1:
            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
          $elif M + 1 == MR:
            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2);
      }
      if (nc & 1) {
        $for M in range(MR):
          $if M % 2 == 1:
            vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
            vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
          $elif M + 1 == MR:
            vst1_lane_s8(c${M}, vout${M}x01234567, 0);
      }

      nc = 0;
    }
  } while (nc != 0);
}