xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8c8-minmax-fp32-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-igemm/c8-neon-mull.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>

#include <arm_neon.h>

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


void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const int8_t** restrict a,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const int8_t* zero,
    const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8 * sizeof(int8_t));
  int8_t* c0 = c;

  do {
    int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
    int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));

    size_t p = ks;
    do {
      const int8_t* restrict a0 = a[0];
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
      }
      a += 1;

      size_t k = kc;
      // 2x partial unrolled loop to load 16 bytes at a time using MLA.
      while (k >= 16 * sizeof(int8_t)) {
        const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
        const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;

        const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));

        const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
        vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
        vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
        const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
        vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
        vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
        const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
        vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
        vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
        const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
        vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
        vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
        const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
        vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
        vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
        const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
        vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
        vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
        const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
        vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
        vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
        const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
        int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
        vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
        vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);

        k -= 16 * sizeof(int8_t);
      }

      // Handle 8 bytes at a time using MUL.
      if (k != 0) {
        const int8x8_t va0 = vld1_s8(a0); a0 += 8;

        const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
        vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
        const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
        vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
        const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
        vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
        const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
        vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
        const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
        vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
        const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
        vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
        const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
        vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
        const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
        const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
        vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);

        k -= 8 * sizeof(int8_t);
      }

      p -= 1 * sizeof(void*);
    } while (p != 0);

#if XNN_ARCH_ARM64
    const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
    const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
    const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
    const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);

    int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
    int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
#else
    const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
    const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
    const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
    const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
    const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
    const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
    int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
    const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
    const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
    const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
    const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
    const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
    const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
    int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
#endif

    float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
    float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);

    const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
    vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
    const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
    vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);

    const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
    vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
    vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));

    const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
    vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
    vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);

#if XNN_ARCH_ARM64
    int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);


    int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
#else
    int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));


    int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
#endif

    const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
    vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);

    const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
    vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);

    if (nc >= 8) {
      vst1_s8(c0 + 0, vout0x01234567);

      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

      a = (const int8_t**restrict) ((uintptr_t) a - ks);

      nc -= 8;
    } else {
      if (nc & 4) {
        vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
        vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
      }
      if (nc & 2) {
        vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
        vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
      }
      if (nc & 1) {
        vst1_lane_s8(c0, vout0x01234567, 0);
      }

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