src/qs8-gemm/gen/1x8c4-minmax-neondot.c

// Auto-generated file. Do not edit!
//   Template: src/qs8-gemm/MRxNRc4-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_qs8_gemm_minmax_ukernel_1x8c4__neondot(
    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 xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN {
  assert(mr != 0);
  assert(mr <= 1);
  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);
  const int8_t* a0 = a;
  int8_t* c0 = c;

  // Loop over groups of 8 columns.
  do {
    // Initialize accumulators with bias. 8 bias values are loaded from the
    // weight matrix, at the start of the group of 8 columns.
    int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
    int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

    // Inner accumulation loop along the 8 columns.
    size_t k = kc;
    // 2x partial unrolled loop to load 8 bytes at a time.
    while (k >= 8 * sizeof(int8_t)) {
      // Load a 1x8 block of activations.
      const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;

      // Load a 8x8 block of weights.
      const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
      const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
      const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
      const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);

      // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
      vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
      vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
      vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
      vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);

      k -= 8 * sizeof(int8_t);
    }
    // Handle up to 4 final positions of `k`
    if XNN_UNLIKELY(k != 0) {
      // Load a 1x4 block of activations.
      const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;

      // Load a 4x8 block of weights.
      const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
      const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);

      // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
      vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
      vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
    }

    // Post-accumulation work
    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 int32x4_t vproduct0x0123 = vqrdmulhq_n_s32(vacc0x0123, params->neon.multiplier);
    const int32x4_t vproduct0x4567 = vqrdmulhq_n_s32(vacc0x4567, params->neon.multiplier);

    vacc0x0123 = vsraq_n_s32(vproduct0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
    vacc0x4567 = vsraq_n_s32(vproduct0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);

    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);

    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
#if XNN_ARCH_ARM64
    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);

    int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
#else
    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);

    int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
#endif
    const int8x8_t voutput_min = vld1_dup_s8(&params->neon.output_min);
    const int8x8_t voutput_max = vld1_dup_s8(&params->neon.output_max);

    vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);

    vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);

    if (nc >= 8) {
      // Main case where there the 8 columns fit in the destination.
      vst1_s8(c0 + 0, vout0x01234567);

      // Advance to the next 8 columns.
      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

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

      nc -= 8;
    } else {
      // Final case where not all of the 8 columns fit in the destination.
      if (nc & 4) {
        vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
        vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
      }
      if (nc & 2) {
        vst1_lane_u16(__builtin_assume_aligned(c0, 1), 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);
}