// 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 NR % 4 == 0
$assert DATATYPE in ["F32", "QC8"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"

#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/gemm.h"


$RANGE_MR = list(reversed(range(MR))) if INC else list(range(MR))
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}s4__${"neonfma" if FMA else "neon"}(
    size_t mr,
    size_t nc,
    size_t kc,
    const float* restrict a,
    size_t a_stride,
    $if DATATYPE == "F32":
      const float* restrict w,
    $else:
      const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const float* restrict acc,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const float* a0 = a;
  float* c0 = c;
  $for M in range(1, MR):
    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
    float* c${M} = (float*) ((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 {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${N//4} = vld1q_f32(acc); acc += 4;
    $else:
      $for N in range(0, NR, 4):
        $if DATATYPE == "F32":
          float32x4_t vacc0x${N//4} = vld1q_f32(w); w += 4;
        $elif DATATYPE == "QC8":
          float32x4_t vacc0x${N//4} = vld1q_f32(w); w = (const float*) w + 4;
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${N//4} = vacc0x${N//4};

    size_t k = kc;
    while (k >= 4 * sizeof(float)) {
      $for M in range(MR):
        float32x4_t va${M} = vld1q_f32(a${M}); a${M} += 4;

      $for L in range(4):

        $if DATATYPE == "F32":
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
        $elif DATATYPE == "QC8":
          $for N in range(0, NR, 8):
            const int8x8_t vw${ABC[N:N+8]}c${L} = vld1_s8(w); w = (const int8_t*) w + 8;
          $for N in range(0, NR, 8):
            const int16x8_t vxw${ABC[N:N+8]}c${L} = vmovl_s8(vw${ABC[N:N+8]}c${L});
          $for N in range(0, NR, 8):
            const int32x4_t vxw${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+8]}c${L}));
            const int32x4_t vxw${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vxw${ABC[N:N+8]}c${L}));
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vxw${ABC[N:N+4]}c${L});

        $for N in range(0, NR, 4):
          $for M in range(MR):
            vacc${M}x${N//4} = ${VMULADDQ_F32}(vacc${M}x${N//4}, va${M}, vb${ABC[N:N+4]}c${L});

        $if L + 1 != 4:
          $for M in range(MR):
            va${M} = vextq_f32(va${M}, va${M}, 1);

      k -= 4 * sizeof(float);
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        float32x4_t va${M} = vld1q_f32(a${M}); a${M} = (const float*) ((uintptr_t) a${M} + k);

      $for L in range(4):

        $if DATATYPE == "F32":
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
        $elif DATATYPE == "QC8":
          $for N in range(0, NR, 8):
            const int8x8_t vw${ABC[N:N+8]}c${L} = vld1_s8(w); w = (const int8_t*) w + 8;
          $for N in range(0, NR, 8):
            const int16x8_t vxw${ABC[N:N+8]}c${L} = vmovl_s8(vw${ABC[N:N+8]}c${L});
          $for N in range(0, NR, 8):
            const int32x4_t vxw${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vxw${ABC[N:N+8]}c${L}));
            const int32x4_t vxw${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vxw${ABC[N:N+8]}c${L}));
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vxw${ABC[N:N+4]}c${L});

        $for N in range(0, NR, 4):
          $for M in range(MR):
            const float32x4_t vmska${M}x${ABC[N:N+4]}c${L} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va${M}), vceqq_f32(vb${ABC[N:N+4]}c${L}, vmovq_n_f32(0.0f))));
            vacc${M}x${N//4} = ${VMULADDQ_F32}(vacc${M}x${N//4}, vmska${M}x${ABC[N:N+4]}c${L}, vb${ABC[N:N+4]}c${L});

        $if L + 1 != 4:
          $for M in range(MR):
            va${M} = vextq_f32(va${M}, va${M}, 1);
    }
    $if DATATYPE == "QC8":
      $for N in range(0, NR, 4):
        const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;
        $for M in range(MR):
          vacc${M}x${N//4} = vmulq_f32(vacc${M}x${N//4}, vscale${ABC[N:N+4]});
    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${N//4} = vminq_f32(vacc${M}x${N//4}, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${N//4} = vmaxq_f32(vacc${M}x${N//4}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in RANGE_MR:
        vst1q_f32(c${M}, vacc${M}x0);
        $for N in range(4, NR, 4):
          vst1q_f32(c${M} + ${N}, vacc${M}x${N//4});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      $for M in RANGE_MR:
        a${M} = (const float*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};

    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 2:
              $for N in range(0, 1 << LOG2N, 4):
                $for M in RANGE_MR:
                  vst1q_f32(c${M}, vacc${M}x${N//4}); c${M} += 4;

              $for M in RANGE_MR:
                $for N in range(0, NR - (1 << LOG2N), 4):
                  vacc${M}x${N//4} = vacc${M}x${(N + (1 << LOG2N))//4};
            $elif LOG2N == 1:
              $for M in RANGE_MR:
                vst1_f32(c${M}, vacc${M}); c${M} += 2;

              $for M in RANGE_MR:
                vacc${M} = vget_high_f32(vacc${M}x0);
            $elif LOG2N == 0:
              $for M in RANGE_MR:
                vst1_lane_f32(c${M}, vacc${M}, 0);
          }
          $if LOG2N == 2:
            $for M in RANGE_MR:
              float32x2_t vacc${M} = vget_low_f32(vacc${M}x0);

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