// Auto-generated file. Do not edit!
//   Template: src/qs8-gemm/c4-neondot.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/gemm.h"
#include "xnnpack/math.h"


void xnn_qd8_f16_qb4w_gemm_minmax_ukernel_2x16c4__neondotfp16arith(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    xnn_float16* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const struct xnn_f16_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
    const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= 2);
  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, 4 * sizeof(int8_t));
  const int8_t* a0 = a;
  uint16_t* c0 = (uint16_t*) c;
  const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
  uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr != 2) {
    a1 = a0;
    c1 = c0;
  }

  size_t bl = params->scalar.blocksize;
  assert(bl <= kc);
  assert(bl != 0);
  assert(bl % 32 == 0);
  const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
  // Loop over groups of 16 columns.
  do {
    // Initialize accumulators with bias. 16 bias values are loaded from the
    // weight matrix, at the start of the group of 16 columns.
    const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
    const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
    float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
    float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
    const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
    float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
    float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
    const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
    float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
    float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
    const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
    float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
    float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);

    for (size_t kb=0; kb < kc; kb += bl) {
      int32x4_t vacc0x0123 = vdupq_n_s32(0);
      int32x4_t vacc1x0123 = vdupq_n_s32(0);
      int32x4_t vacc0x4567 = vdupq_n_s32(0);
      int32x4_t vacc1x4567 = vdupq_n_s32(0);
      int32x4_t vacc0x89AB = vdupq_n_s32(0);
      int32x4_t vacc1x89AB = vdupq_n_s32(0);
      int32x4_t vacc0xCDEF = vdupq_n_s32(0);
      int32x4_t vacc1xCDEF = vdupq_n_s32(0);
      // Inner accumulation loop along the 16 columns.
      size_t k = bl;
      // 2x partial unrolled loop to load 8 bytes at a time.
      while (k >= 8 * sizeof(int8_t)) {
        // Load a 2x8 block of activations.
        const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
        const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;

        // Load a 8x16 block of weights.
        const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
        const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
        const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
        const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
        const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
        const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
        const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
        const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);

        // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
        vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
        vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
        vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
        vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
        vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
        vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
        vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
        vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
        vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
        vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
        vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
        vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
        vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
        vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
        vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
        vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);

        k -= 8 * sizeof(int8_t);
      }
      // Handle up to 4 final positions of `k`
      if XNN_UNLIKELY(k != 0) {
        // Load a 2x4 block of activations.
        const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
        const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;

        // Load a 4x16 block of weights.
        const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
        const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;

        // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
        vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
        vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
        vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
        vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
        vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
        vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
        vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
        vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
      }
      const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
      const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
      const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
      const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;

      float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
      float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
      vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
      vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
      float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
      float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
      vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
      vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
      float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
      float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
      vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
      vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
      float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
      float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
      vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
      vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
    }

    const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
    vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
    vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
    vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
    vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
    vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
    vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
    vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
    vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);


    const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
    vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
    vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
    const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
    vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
    vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
    const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
    vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
    vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
    const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
    vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
    vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);

    float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
    float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
    float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
    float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));

    const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
    vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
    vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
    vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
    vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
    const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
    vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
    vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
    vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
    vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
    if XNN_LIKELY(nc >= 16) {
      vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
      vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
      vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
      vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));

      a0 = (const int8_t*) ((uintptr_t) a0 - kc);
      a1 = (const int8_t*) ((uintptr_t) a1 - kc);

      c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
      c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);

      nc -= 16;
    } else {
     if (nc & 8) {
       vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
       vfp16out0x01234567 = vfp16out0x89ABCDEF;
       vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
       vfp16out1x01234567 = vfp16out1x89ABCDEF;
     }
     float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
     float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
     if (nc & 4) {
       vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
       vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
       vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
       vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
     }
     if (nc & 2) {
       vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
       vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
       vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
       vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
     }
     if (nc & 1) {
       vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
       vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
     }
      nc = 0;
    }
  } while (nc != 0);
}