// Copyright 2019 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 MR % 8 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/prefetch.h"
#include "xnnpack/spmm.h"


void xnn_f16_spmm_minmax_ukernel_${MR}x${NR}__neonfp16arith${"_x%d" % UNROLL if UNROLL > 1 else ""}(
    size_t mc,
    size_t nc,
    const xnn_float16* input,
    const xnn_float16* weights,
    const int32_t* widx_dmap,
    const uint32_t* nidx_nnzmap,
    xnn_float16* output,
    size_t output_stride,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mc != 0);
  assert(mc % sizeof(uint16_t) == 0);
  assert(nc != 0);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;

  #if XNN_ARCH_ARM64
    const uint16x8x2_t vminmax = vld2q_dup_u16((const uint16_t*) &params->scalar.min);
    const float16x8_t vmin = vreinterpretq_f16_u16(vminmax.val[0]);
    const float16x8_t vmax = vreinterpretq_f16_u16(vminmax.val[1]);
  #else
    // vld2_dup is to work around aarch32 clang bug with vld1q_dup
    const uint16x4x2_t vminmax = vld2_dup_u16((const uint16_t*) &params->scalar.min);
    const float16x8_t vmin = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[0], vminmax.val[0]));
    const float16x8_t vmax = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[1], vminmax.val[1]));
  #endif

  size_t output_decrement = output_stride * nc - ${MR} * sizeof(uint16_t);
  while XNN_LIKELY(mc >= ${MR} * sizeof(uint16_t)) {
    const uint16_t* w = (const uint16_t*) weights;
    const int32_t* dmap = widx_dmap;
    const uint32_t* nnzmap = nidx_nnzmap;
    size_t n = nc;
    do {
      uint32_t nnz = *nnzmap++;
      $if UNROLL > 1:
        float16x8_t vacc01234567x0 = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
        $for K in range(1, UNROLL):
          float16x8_t vacc01234567x${K} = vreinterpretq_f16_u16(vmovq_n_u16(0));
        $for M in range(8, MR, 8):
          float16x8_t vacc${ABC[M:M+8]}x0 = vacc01234567x0;
          $for K in range(1, UNROLL):
            float16x8_t vacc${ABC[M:M+8]}x${K} = vreinterpretq_f16_u16(vmovq_n_u16(0));
        for (; nnz >= ${UNROLL}; nnz -= ${UNROLL}) {
          $for K in range(UNROLL):
            const intptr_t diff${K} = dmap[${K}];
          dmap += ${UNROLL};
          $for K in range(UNROLL):
            const float16x8_t va01234567x${K} = vreinterpretq_f16_u16(vld1q_u16(i));
            $for M in range(8, MR, 8):
              const float16x8_t va${ABC[M:M+8]}x${K} = vreinterpretq_f16_u16(vld1q_u16(i + ${M}));
            i = (const uint16_t*) ((uintptr_t) i + (uintptr_t) diff${K});
            $for M in range(0, MR, 32):
              xnn_prefetch_to_l1(i + ${M+32});
            const float16x8_t vw${K} = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
            xnn_prefetch_to_l1(w + 64);
            $for M in range(0, MR, 8):
              vacc${ABC[M:M+8]}x${K} = vfmaq_f16(vacc${ABC[M:M+8]}x${K}, va${ABC[M:M+8]}x${K}, vw${K});
        }
        $for M in range(0, MR, 8):
          float16x8_t vacc${ABC[M:M+8]} = vacc${ABC[M:M+8]}x0;
        $for K in range(1, UNROLL):
          $for M in range(0, MR, 8):
            vacc${ABC[M:M+8]} = vaddq_f16(vacc${ABC[M:M+8]}, vacc${ABC[M:M+8]}x${K});
      $else:
        float16x8_t vacc01234567 = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
        $for M in range(8, MR, 8):
          float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
      if XNN_LIKELY(nnz != 0) {
        do {
          const intptr_t diff = *dmap++;
          const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(i));
          $for M in range(8, MR, 8):
            const float16x8_t va${ABC[M:M+8]} = vreinterpretq_f16_u16(vld1q_u16(i + ${M}));
          i = (const uint16_t*) ((uintptr_t) i + (uintptr_t) diff);
          $for M in range(0, MR, 32):
            xnn_prefetch_to_l1(i + ${M+32});
          const float16x8_t vw = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
          xnn_prefetch_to_l1(w + 64);
          $for M in range(0, MR, 8):
            vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vw);
        } while (--nnz != 0);
      }
      $for M in range(0, MR, 8):
        float16x8_t vout${ABC[M:M+8]} = vminq_f16(vacc${ABC[M:M+8]}, vmax);
      $for M in range(0, MR, 8):
        vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
      vst1q_u16(o, vreinterpretq_u16_f16(vout01234567));
      $for M in range(8, MR, 8):
        vst1q_u16(o + ${M}, vreinterpretq_u16_f16(vout${ABC[M:M+8]}));
      o = (uint16_t*) ((uintptr_t) o + output_stride);
    } while (--n != 0);
    o = (uint16_t*) ((uintptr_t) o - output_decrement);
    i += ${MR};
    mc -= ${MR} * sizeof(uint16_t);
  }
  if XNN_UNLIKELY(mc != 0) {
    $for LOG2M in reversed(range((MR - 1).bit_length())):
      $SUBMR = 1 << LOG2M
      $if SUBMR * 2 >= MR:
        output_decrement += ${MR - SUBMR} * sizeof(uint16_t);
      $else:
        output_decrement += ${SUBMR} * sizeof(uint16_t);
      if (mc & (${SUBMR} * sizeof(uint16_t))) {
        const uint16_t* w = (const uint16_t*) weights;
        const int32_t* dmap = widx_dmap;
        const uint32_t* nnzmap = nidx_nnzmap;
        size_t n = nc;
        do {
          uint32_t nnz = *nnzmap++;
          $if SUBMR <= 4:
            float16x4_t vacc${ABC[0:SUBMR]} = vreinterpret_f16_u16(vld1_dup_u16(w)); w += 1;
          $else:
            float16x8_t vacc01234567 = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
          $for M in range(8, SUBMR, 8):
            float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
          if XNN_LIKELY(nnz != 0) {
            do {
              const intptr_t diff = *dmap++;
              $if SUBMR == 1:
                const float16x4_t va0 = vreinterpret_f16_u16(vld1_dup_u16(i));
              $elif SUBMR == 2:
                const float16x4_t va01 = vreinterpret_f16_u32(vld1_dup_u32((const void*) i));
              $elif SUBMR == 4:
                const float16x4_t va0123 = vreinterpret_f16_u16(vld1_u16(i));
              $else:
                const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(i));
              $for M in range(8, SUBMR, 8):
                const float16x8_t va${ABC[M:M+8]} = vreinterpretq_f16_u16(vld1q_u16(i + ${M}));
              i = (const uint16_t*) ((uintptr_t) i + (uintptr_t) diff);
              $if SUBMR <= 4:
                const float16x4_t vw = vreinterpret_f16_u16(vld1_dup_u16(w)); w += 1;
              $else:
                const float16x8_t vw = vreinterpretq_f16_u16(vld1q_dup_u16(w)); w += 1;
              $if SUBMR <= 4:
                vacc${ABC[0:SUBMR]} = vfma_f16(vacc${ABC[0:SUBMR]}, va${ABC[0:SUBMR]}, vw);
              $else:
                $for M in range(0, SUBMR, 8):
                  vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vw);
            } while (--nnz != 0);
          }
          $if SUBMR <= 4:
            float16x4_t vout${ABC[0:SUBMR]} = vmin_f16(vacc${ABC[0:SUBMR]}, vget_low_f16(vmax));
            vout${ABC[0:SUBMR]} = vmax_f16(vout${ABC[0:SUBMR]}, vget_low_f16(vmin));
            $if SUBMR == 1:
              vst1_lane_u16(o, vreinterpret_u16_f16(vout${ABC[0]}), 0);
            $elif SUBMR == 2:
              vst1_lane_u32((void*) o, vreinterpret_u32_f16(vout${ABC[0:SUBMR]}), 0);
            $else:
              vst1_u16(o, vreinterpret_u16_f16(vout${ABC[0:SUBMR]}));
          $else:
            $for M in range(0, SUBMR, 8):
              float16x8_t vout${ABC[M:M+8]} = vminq_f16(vacc${ABC[M:M+8]}, vmax);
            $for M in range(0, SUBMR, 8):
              vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
            vst1q_u16(o, vreinterpretq_u16_f16(vout01234567));
            $for M in range(8, SUBMR, 8):
              vst1q_u16(o + ${M}, vreinterpretq_u16_f16(vout${ABC[M:M+8]}));
          o = (uint16_t*) ((uintptr_t) o + output_stride);
        } while (--n != 0);
        o = (uint16_t*) ((uintptr_t) o - output_decrement);
        i += ${SUBMR};
      }
  }
}