// 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 DATATYPE in ["F32", "QC8"]
$assert NR == 2
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/gemm.h>


$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${"aarch64_neonfma" if FMA else "neon"}_lane_ld64(
    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:
      $for M in range(0, MR):
        float32x2_t vacc${M}x01 = vld1_f32(w); w += 2;
    $else:
      $if DATATYPE == "F32":
        float32x2_t vacc0x01 = vld1_f32(w); w += 2;
      $else:
        float32x2_t vacc0x01 = vreinterpret_f32_u8(vld1_u8(w)); w = (const float*) w + 2;
      $for M in range(1, MR):
        float32x2_t vacc${M}x01 = vacc0x01;

    size_t k = kc;
    for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
      $for M in range(MR):
        const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2;

      $if DATATYPE == "F32":
        const float32x4_t vb01c01 = vld1q_f32(w); w += 4;
      $elif DATATYPE == "QC8":
        const uint32x2_t vtmpb = vld1_dup_u32(w); w = (const int8_t*) w + 4;
        const int32x4_t vtmpi = vmovl_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_u32(vtmpb))));
        const float32x4_t vb01c01 = vcvtq_f32_s32(vtmpi);
      const float32x2_t vb01c0 = vget_low_f32(vb01c01);
      const float32x2_t vb01c1 = vget_high_f32(vb01c01);
      $for L in range(2):
        $if FMA:
          #if XNN_ARCH_ARM64
            $for M in range(MR):
              vacc${M}x01 = vfma_lane_f32(vacc${M}x01, vb01c${L}, va${M}, ${L});
          #else
            $for M in range(MR):
              const float32x2_t va${M}c${L} = vdup_lane_f32(va${M}, ${L});
            $for M in range(MR):
              vacc${M}x01 = vfma_f32(vacc${M}x01, va${M}c${L}, vb01c${L});
          #endif
        $else:
          $for M in range(MR):
            vacc${M}x01 = vmla_lane_f32(vacc${M}x01, vb01c${L}, va${M}, ${L});
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const float32x2_t va${M} = vld1_dup_f32(a${M}); a${M} += 1;

      $if DATATYPE == "F32":
        const float32x2_t vb01 = vld1_f32(w); w += 2;
      $elif DATATYPE == "QC8":
        const uint16x4_t vtmpb = vld1_dup_u16(w); w = (const int8_t*) w + 2;
        const int32x2_t vtmpi = vget_low_s32(vmovl_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_u16(vtmpb)))));
        const float32x2_t vb01 = vcvt_f32_s32(vtmpi);

      $for M in range(MR):
        $if FMA:
          vacc${M}x01 = vfma_f32(vacc${M}x01, va${M}, vb01);
        $else:
          vacc${M}x01 = vmla_f32(vacc${M}x01, va${M}, vb01);
    }

    $if DATATYPE == "QC8":
      const float32x2_t vscale = vreinterpret_f32_u8(vld1_u8(w)); w = (const float*) w + 2;
      $for M in range(MR):
        vacc${M}x01 = vmul_f32(vacc${M}x01, vscale);
    const float32x2_t vmax = vld1_dup_f32(&params->scalar.max);
    $for M in range(MR):
      vacc${M}x01 = vmin_f32(vacc${M}x01, vmax);
    const float32x2_t vmin = vld1_dup_f32(&params->scalar.min);
    $for M in range(MR):
      vacc${M}x01 = vmax_f32(vacc${M}x01, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        vst1_f32(c${M}, vacc${M}x01);
        c${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 {
      assert(nc == 1);
      $for M in range(MR):
        vst1_lane_f32(c${M}, vacc${M}x01, 0);

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