// 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", "QC4", "QC8"]
$assert NR == 2
$assert MR % 2 == 0
#include <assert.h>

$if DATATYPE in ["QC8", "QC4"]:
  #include <smmintrin.h>
$else:
  #include <xmmintrin.h>

#include <xnnpack/gemm.h>
$if DATATYPE == "QC8":
  #include <xnnpack/unaligned.h>


$if DATATYPE in ["QC8", "QC4"]:
  $ISA = {2: "sse2", 4: "sse41"}[SSE]
$else:
  $ISA = "sse"
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm_minmax_ukernel_${MR}x${NR}c4__${ISA}(
    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 DATATYPE == "QC4":
      const union xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
    $else:
      const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  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);

  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 {
    __m128 vacc0x0c4 = _mm_load_ss(w);
    $for N in range(1, NR):
      $if DATATYPE == "F32":
        __m128 vacc0x${N}c4 = _mm_load_ss(w + ${N});
      $else:
        __m128 vacc0x${N}c4 = _mm_load_ss((const float*) w + ${N});
    $for M in range(1, MR):
      $for N in range(NR):
        __m128 vacc${M}x${N}c4 = vacc0x${N}c4;
    $if DATATYPE == "F32":
      w += ${NR};
    $else:
      w = (const float*) w + ${NR};

    size_t k = kc;
    for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
      $for M in range(MR):
        const __m128 va${M} = _mm_loadu_ps(a${M});
        a${M} += 4;

      $if DATATYPE == "F32":
        const __m128 vb0 = _mm_loadu_ps(w);
        $for N in range(1, NR):
          const __m128 vb${N} = _mm_loadu_ps(w + ${N * 4});
        w += ${NR * 4};
      $else:
        $if SSE >= 4:
          const __m128i vbi0 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_u32(w)));
          $for N in range(1, NR):
            const __m128i vbi${N} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const int8_t*) w + ${N * 4})));
          $for N in range(NR):
            const __m128 vb${N} = _mm_cvtepi32_ps(vbi${N});
        $else:
          $for N in range(0, NR, 2):
            const __m128i vb${N}${N+1} = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 4}));
          $for N in range(0, NR, 2):
            const __m128i vbw${N}${N+1} = _mm_unpacklo_epi8(vb${N}${N+1}, vb${N}${N+1});
            const __m128 vb${N} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(vbw${N}${N+1}, vbw${N}${N+1}), 24));
            const __m128 vb${N+1} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(vbw${N}${N+1}, vbw${N}${N+1}), 24));
        w = (const int8_t*) w + ${NR * 4};

      $for M in range(MR):
        $for N in range(NR):
          vacc${M}x${N}c4 = _mm_add_ps(vacc${M}x${N}c4, _mm_mul_ps(va${M}, vb${N}));
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const __m128 va${M} = _mm_loadu_ps(a${M});
        a${M} = (const float*) ((uintptr_t) a${M} + k);

      $if DATATYPE == "F32":
        const __m128 vb0 = _mm_loadu_ps(w);
        $for N in range(1, NR):
          const __m128 vb${N} = _mm_loadu_ps(w + ${N * 4});
        w += ${NR * 4};
      $else:
        $if SSE >= 4:
          const __m128i vbi0 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_u32(w)));
          $for N in range(1, NR):
            const __m128i vbi${N} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const int8_t*) w + ${N * 4})));
          $for N in range(NR):
            const __m128 vb${N} = _mm_cvtepi32_ps(vbi${N});
        $else:
          $for N in range(0, NR, 2):
            const __m128i vb${N}${N+1} = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 4}));
          $for N in range(0, NR, 2):
            const __m128i vbw${N}${N+1} = _mm_unpacklo_epi8(vb${N}${N+1}, vb${N}${N+1});
            const __m128 vb${N} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(vbw${N}${N+1}, vbw${N}${N+1}), 24));
            const __m128 vb${N+1} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(vbw${N}${N+1}, vbw${N}${N+1}), 24));
        w = (const int8_t*) w + ${NR * 4};

      $for N in range(NR):
        const __m128 vmask${N} = _mm_cmpeq_ps(_mm_setzero_ps(), vb${N});

      $for M in range(MR):
        $for N in range(NR):
          vacc${M}x${N}c4 = _mm_add_ps(vacc${M}x${N}c4, _mm_mul_ps(_mm_andnot_ps(vmask${N}, va${M}), vb${N}));
    }

    $for M in range(MR):
      const __m128 vacc${M}x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc${M}x0c4, vacc${M}x1c4), _mm_unpackhi_ps(vacc${M}x0c4, vacc${M}x1c4));

    $for M in range(0, MR, 2):
      __m128 vacc${M}${M+1}x01 = _mm_add_ps(_mm_movelh_ps(vacc${M}x01c2, vacc${M+1}x01c2), _mm_movehl_ps(vacc${M+1}x01c2, vacc${M}x01c2));

    $if DATATYPE in ["QC8", "QC4"]:
      const __m128 vscalex01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w));
      const __m128 vscale2x01 = _mm_movelh_ps(vscalex01, vscalex01);
      w = (const float*) w + 2;
      $for M in range(0, MR, 2):
        vacc${M}${M+1}x01 = _mm_mul_ps(vacc${M}${M+1}x01, vscale2x01);
    const __m128 vmax = _mm_load_ps(params->sse.max);
    $for M in range(0, MR, 2):
      vacc${M}${M+1}x01 = _mm_min_ps(vacc${M}${M+1}x01, vmax);

    const __m128 vmin = _mm_load_ps(params->sse.min);
    $for M in range(0, MR, 2):
      vacc${M}${M+1}x01 = _mm_max_ps(vacc${M}${M+1}x01, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(0, MR, 2)):
        _mm_storel_pi((__m64*) c${M}, vacc${M}${M+1}x01);
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
        a${M} = (const float*) ((uintptr_t) a${M} - kc);
        _mm_storeh_pi((__m64*) c${M+1}, vacc${M}${M+1}x01);
        c${M+1} = (float*) ((uintptr_t) c${M+1} + cn_stride);
        a${M+1} = (const float*) ((uintptr_t) a${M+1} - kc);

      nc -= ${NR};
    } else {
      assert(nc == 1);
      $for M in reversed(range(0, MR, 2)):
        _mm_store_ss(c${M}, vacc${M}${M+1}x01);
        _mm_store_ss(c${M+1}, _mm_movehl_ps(vacc${M}${M+1}x01, vacc${M}${M+1}x01));

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