// Copyright 2023 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
$assert NR % 4 == 0
$assert SR == 4
$assert KBLOCK == 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"

#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include <immintrin.h>

#include "xnnpack/packw.h"
$if PREFETCH:
  #include "xnnpack/prefetch.h"


void xnn_x32_packw_gemm_goi_ukernel_x${NR}s4__avx_u${KBLOCK}${"_prfm" if PREFETCH else ""}(
  size_t g,
  size_t nc,
  size_t kc,
  size_t nr,
  size_t kr,
  size_t sr,
  const uint32_t* weights,
  const uint32_t* bias,
  const void* scale,
  uint32_t* packed_weights,
  size_t extra_bytes,
  const void* params)
{
  assert(g != 0);
  assert(nc != 0);
  assert(kc != 0);
  assert(nr == ${NR});
  assert(kr == 1);
  assert(sr == ${SR});
  assert(weights != NULL);
  assert(packed_weights != NULL);

  const float* b = (const float*) bias;
  float* packed_w = (float*) packed_weights;
  do {
    // NC main loop multiple of ${NR}
    const float* w0 = (const float*) weights;
    size_t n = nc;

    for (; n >= ${NR}; n -= ${NR}) {
      if XNN_LIKELY(b != NULL) {
        const __m256 vb0 = _mm256_loadu_ps(b);
        $for N in range(8,NR,8):
          const __m256 vb${N} = _mm256_loadu_ps(b + ${N});
        _mm256_store_ps(packed_w, vb0);
        $for N in range(8,NR,8):
          _mm256_store_ps(packed_w + ${N}, vb${N});
        b += ${NR};
      } else {
        const __m256 vzero = _mm256_setzero_ps();
        _mm256_store_ps(packed_w, vzero);
        $for N in range(8,NR,8):
          _mm256_store_ps(packed_w + ${N}, vzero);
      }
      packed_w += ${NR};

      $for N in range(1, NR):
        const float* w${N} = w${N-1} + kc;
      $if PREFETCH:
        $for N in range(0, NR):
          xnn_prefetch_to_l1((const int8_t*) w${N});
          xnn_prefetch_to_l1((const int8_t*) w${N} + 64);

      // KC main loop multiple of ${KBLOCK}
      size_t k = kc;
      for (; k >= ${KBLOCK}; k -= ${KBLOCK}) {
        // Read blocks of 4x4
        // a b c d
        // e f g h
        // i j k l
        // m n o p
        // Load first 4 rows of N into low part of each register
        $for N in range(0,NR,8):
          $for L in range(4):
            __m256 v${N+L}x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w${N+L}));
            $for K in range(4,KBLOCK,4):
              __m256 v${N+L}x${ABC[K:K+4]} = _mm256_castps128_ps256(_mm_loadu_ps(w${N+L} + ${K}));
            w${N+L} += ${KBLOCK};
        // Load next 4 rows of N into the high part of each register
        $for N in range(0,NR,8):
          $for L in range(4):
            v${N+L}x0123 = _mm256_insertf128_ps(v${N+L}x0123, _mm_loadu_ps(w${N+L+4}), 1);
            $for K in range(4,KBLOCK,4):
              v${N+L}x${ABC[K:K+4]} = _mm256_insertf128_ps(v${N+L}x${ABC[K:K+4]}, _mm_loadu_ps(w${N+L+4} + ${K})), 1);
            w${N+L+4} += ${KBLOCK};

        // Apply SR4 shuffle
        $for N in range(0,NR,8):
          $for L in range(1,4):
            $for K in range(0,KBLOCK,4):
              v${N+L}x${ABC[K:K+4]} = _mm256_permute_ps(v${N+L}x${ABC[K:K+4]}, _MM_SHUFFLE(${(L+3)&3}, ${(L+2)&3}, ${(L+1)&3}, ${L}));

        // Transpose 2x2
        $for N in range(0,NR,8):
          $for K in range(0,KBLOCK,4):
            const __m256 vtmp${N+0}x${ABC[K:K+4]} = _mm256_unpacklo_ps(v${N+0}x${ABC[K:K+4]}, v${N+1}x${ABC[K:K+4]});  // a e b f   from row 0, 1
            const __m256 vtmp${N+1}x${ABC[K:K+4]} = _mm256_unpacklo_ps(v${N+2}x${ABC[K:K+4]}, v${N+3}x${ABC[K:K+4]});  // i m j n   from row 2, 3
            const __m256 vtmp${N+2}x${ABC[K:K+4]} = _mm256_unpackhi_ps(v${N+0}x${ABC[K:K+4]}, v${N+1}x${ABC[K:K+4]});  // c g d h   from row 0, 1
            const __m256 vtmp${N+3}x${ABC[K:K+4]} = _mm256_unpackhi_ps(v${N+2}x${ABC[K:K+4]}, v${N+3}x${ABC[K:K+4]});  // k o l p   from row 2, 3
        $if PREFETCH:
          $for N in range(0, NR):
            xnn_prefetch_to_l1((const int8_t*) w${N} + 128);
        // Transpose 4x4
        $for N in range(0,NR,8):
          $for K in range(0,KBLOCK,4):
            v${N+0}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp${N+0}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+1}x${ABC[K:K+4]})));  // a e i m   from row 0, 1
            v${N+1}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp${N+0}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+1}x${ABC[K:K+4]})));  // b f j n   from row 0, 1
            v${N+2}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp${N+2}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+3}x${ABC[K:K+4]})));  // c g k o   from row 2, 3
            v${N+3}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp${N+2}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+3}x${ABC[K:K+4]})));  // d h l p   from row 2, 3

        _mm256_store_ps(packed_w, v0x0123);
        $for K in range(KBLOCK):
          $for N in range(0,NR,8):
            $if N != 0 or K != 0:
              _mm256_store_ps(packed_w + ${N+K*NR}, v${N+K}x${ABC[K//4*4:K//4*4+4]});
        packed_w += ${NR*KBLOCK};
      }

      // KC remainder (1..3)
      if XNN_UNLIKELY(k != 0) {
        assert(k >= 1);
        assert(k <= 3);
        $for N in range(NR):
          __m128 v${N} = _mm_undefined_ps();

        switch (k) {
          case 1:
            // Read blocks of 4x1
            // a
            // e
            // i
            // m
            $for N in range(NR):
              v${N} = _mm_load_ss(w${N});
              w${N} += 1;
            break;
          case 2:
            // Read blocks of 4x2
            // a b
            // e f
            // i j
            // m n
            $for N in range(NR):
              v${N} = _mm_castpd_ps(_mm_load_sd((const double*) w${N}));
              w${N} += 2;
            break;
          case 3:
          {
            // Read blocks of 4x3
            // a b c
            // e f g
            // i j k
            // m n o
            $for N in range(NR):
              const __m128 v${N}lo = _mm_castpd_ps(_mm_load_sd((const double*) w${N}));
              const __m128 v${N}hi = _mm_load_ss(w${N} + 2);
              v${N} = _mm_movelh_ps(v${N}lo, v${N}hi);
              w${N} += 3;
            break;
          }
          default:
            XNN_UNREACHABLE;
        }

        // Apply SR4 shuffle
        $for N in range(0,NR,4):
          $for L in range(1,4):
            v${N+L} = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v${N+L}), _MM_SHUFFLE(${(L+3)&3}, ${(L+2)&3}, ${(L+1)&3}, ${L})));
        // Transpose 2x2
        $for N in range(0,NR,4):
          const __m128 vtmp${N+0} = _mm_unpacklo_ps(v${N+0}, v${N+1});  // a e b f   from row 0, 1
          const __m128 vtmp${N+1} = _mm_unpacklo_ps(v${N+2}, v${N+3});  // i m j n   from row 2, 3
          const __m128 vtmp${N+2} = _mm_unpackhi_ps(v${N+0}, v${N+1});  // c g d h   from row 0, 1
          const __m128 vtmp${N+3} = _mm_unpackhi_ps(v${N+2}, v${N+3});  // k o l p   from row 2, 3
        // Transpose 4x4
        $for N in range(0,NR,4):
          v${N+0} = _mm_movelh_ps(vtmp${N+0}, vtmp${N+1});  // a e i m   from row 0, 1
          v${N+1} = _mm_movehl_ps(vtmp${N+1}, vtmp${N+0});  // b f j n   from row 0, 1
          v${N+2} = _mm_movelh_ps(vtmp${N+2}, vtmp${N+3});  // c g k o   from row 2, 3
          v${N+3} = _mm_movehl_ps(vtmp${N+3}, vtmp${N+2});  // d h l p   from row 2, 3
        _mm_store_ps(packed_w, v0);
        $for L in range(4):
          $for N in range(0,NR,4):
            $if (N+L) != 0:
              _mm_store_ps(packed_w + ${N+L*NR}, v${N+L});
        packed_w += ${NR*4};
      }
      packed_w = (float*) ((uintptr_t) packed_w + extra_bytes);
      w0 = w${NR-1};
    }

    // NC remainder (1..${NR-1})
    if XNN_UNLIKELY(n != 0) {
      assert(n >= 1);
      assert(n <= ${NR-1});
      if XNN_LIKELY(b != NULL) {
        size_t nb = n;
        do {
          *packed_w++  = *b++;
        } while (--nb != 0);
        packed_w += (${NR} - n);
      } else {
        const __m128 vzero = _mm_setzero_ps();
        _mm_store_ps(packed_w, vzero);
        $for N in range(4, NR, 4):
          _mm_store_ps(packed_w + ${N}, vzero);
        packed_w += ${NR};
      }

      // NR remainder has less than ${NR} rows so last row is not loaded
      $for N in range(1, NR-1):
        const float* w${N} = w${N-1} + kc;
        $if N % 2 == 0:
          if XNN_UNPREDICTABLE(n <= ${N}) {
            w${N} = w${N-1};
          }
        $else:
          if XNN_UNPREDICTABLE(n < ${N+1}) {
            w${N} = w${N-1};
          }

      // KC main loop multiple of ${KBLOCK}
      size_t k = kc;
      for (; k >= ${KBLOCK}; k -= ${KBLOCK}) {
        // Read blocks of 4x4
        // a b c d
        // e f g h
        // i j k l
        // m n o p
        // Load first 4 rows of N into low part of each register
        $for N in range(0,NR,8):
          $for L in range(4):
            __m256 v${N+L}x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w${N+L}));
            $for K in range(4,KBLOCK,4):
              __m256 v${N+L}x${ABC[K:K+4]} = _mm256_castps128_ps256(_mm_loadu_ps(w${N+L} + ${K}));
            w${N+L} += ${KBLOCK};
        // Load next 4 rows of N into the high part of each register
        $for N in range(0,NR,8):
          $for L in range(4):
            $if (N+L+4)<(NR-1):
              v${N+L}x0123 = _mm256_insertf128_ps(v${N+L}x0123, _mm_loadu_ps(w${N+L+4}), 1);
              $for K in range(4,KBLOCK,4):
                v${N+L}x${ABC[K:K+4]} = _mm256_insertf128_ps(v${N+L}x${ABC[K:K+4]}, _mm_loadu_ps(w${N+L+4} + ${K})), 1);
              w${N+L+4} += ${KBLOCK};
        // Apply SR4 shuffle
        $for N in range(0,NR,8):
          $for L in range(1,4):
            $for K in range(0,KBLOCK,4):
              v${N+L}x${ABC[K:K+4]} = _mm256_permute_ps(v${N+L}x${ABC[K:K+4]}, _MM_SHUFFLE(${(L+3)&3}, ${(L+2)&3}, ${(L+1)&3}, ${L}));
        // Transpose 2x2
        $for N in range(0,NR,8):
          $for K in range(0,KBLOCK,4):
            const __m256 vtmp${N+0}x${ABC[K:K+4]} = _mm256_unpacklo_ps(v${N+0}x${ABC[K:K+4]}, v${N+1}x${ABC[K:K+4]});  // a e b f   from row 0, 1
            const __m256 vtmp${N+1}x${ABC[K:K+4]} = _mm256_unpacklo_ps(v${N+2}x${ABC[K:K+4]}, v${N+3}x${ABC[K:K+4]});  // i m j n   from row 2, 3
            const __m256 vtmp${N+2}x${ABC[K:K+4]} = _mm256_unpackhi_ps(v${N+0}x${ABC[K:K+4]}, v${N+1}x${ABC[K:K+4]});  // c g d h   from row 0, 1
            const __m256 vtmp${N+3}x${ABC[K:K+4]} = _mm256_unpackhi_ps(v${N+2}x${ABC[K:K+4]}, v${N+3}x${ABC[K:K+4]});  // k o l p   from row 2, 3
        $if PREFETCH:
          $for N in range(0, NR-1):
            xnn_prefetch_to_l1((const int8_t*) w${N} + 128);
        // Transpose 4x4
        $for N in range(0,NR,8):
          $for K in range(0,KBLOCK,4):
            v${N+0}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp${N+0}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+1}x${ABC[K:K+4]})));  // a e i m   from row 0, 1
            v${N+1}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp${N+0}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+1}x${ABC[K:K+4]})));  // b f j n   from row 0, 1
            v${N+2}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp${N+2}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+3}x${ABC[K:K+4]})));  // c g k o   from row 2, 3
            v${N+3}x${ABC[K:K+4]} = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp${N+2}x${ABC[K:K+4]}), _mm256_castps_pd(vtmp${N+3}x${ABC[K:K+4]})));  // d h l p   from row 2, 3

        _mm256_store_ps(packed_w, v0x0123);
        $for K in range(KBLOCK):
          $for N in range(0,NR,8):
            $if N != 0 or K != 0:
              _mm256_store_ps(packed_w + ${N+K*NR}, v${N+K}x${ABC[K//4*4:K//4*4+4]});
        packed_w += ${NR*KBLOCK};
      }

      // KC remainder (1..3)
      if XNN_UNLIKELY(k != 0) {
        assert(k >= 1);
        assert(k <= 3);
        $for N in range(NR):
          __m128 v${N} = _mm_undefined_ps();

        switch (k) {
          case 1:
            // Read blocks of 4x1
            // a
            // e
            // i
            // m
            $for N in range(NR-1):
              v${N} = _mm_load_ss(w${N});
              w${N} += 1;
            break;
          case 2:
            // Read blocks of 4x2
            // a b
            // e f
            // i j
            // m n
            $for N in range(NR-1):
              v${N} = _mm_castpd_ps(_mm_load_sd((const double*) w${N}));
              w${N} += 2;
            break;
          case 3:
          {
            // Read blocks of 4x3
            // a b c
            // e f g
            // i j k
            // m n o
            $for N in range(NR-1):
              const __m128 v${N}lo = _mm_castpd_ps(_mm_load_sd((const double*) w${N}));
              const __m128 v${N}hi = _mm_load_ss(w${N} + 2);
              v${N} = _mm_movelh_ps(v${N}lo, v${N}hi);
              w${N} += 3;
            break;
          }
          default:
            XNN_UNREACHABLE;
        }

        // Apply SR4 shuffle
        $for N in range(0,NR,4):
          $for L in range(1,4):
            $if (N+L) != (NR-1):
              v${N+L} = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v${N+L}), _MM_SHUFFLE(${(L+3)&3}, ${(L+2)&3}, ${(L+1)&3}, ${L})));
        // Transpose 2x2
        $for N in range(0,NR,4):
          $for K in range(0,KBLOCK,4):
            const __m128 vtmp${N+0} = _mm_unpacklo_ps(v${N+0}, v${N+1});  // a e b f   from row 0, 1
            const __m128 vtmp${N+1} = _mm_unpacklo_ps(v${N+2}, v${min(N+3, NR-2)});  // i m j n   from row 2, 3
            const __m128 vtmp${N+2} = _mm_unpackhi_ps(v${N+0}, v${N+1});  // c g d h   from row 0, 1
            const __m128 vtmp${N+3} = _mm_unpackhi_ps(v${N+2}, v${min(N+3, NR-2)});  // k o l p   from row 2, 3
        // Transpose 4x4
        $for N in range(0,NR,4):
          $for K in range(0,KBLOCK,4):
            v${N+0} = _mm_movelh_ps(vtmp${N+0}, vtmp${N+1});  // a e i m   from row 0, 1
            v${N+1} = _mm_movehl_ps(vtmp${N+1}, vtmp${N+0});  // b f j n   from row 0, 1
            v${N+2} = _mm_movelh_ps(vtmp${N+2}, vtmp${N+3});  // c g k o   from row 2, 3
            v${N+3} = _mm_movehl_ps(vtmp${N+3}, vtmp${N+2});  // d h l p   from row 2, 3
        _mm_store_ps(packed_w, v0);
        $for L in range(4):
          $for N in range(0,NR,4):
            $if (N+L) != 0:
              _mm_store_ps(packed_w + ${N+L*NR}, v${N+L});
        packed_w += ${NR*4};
      }
      packed_w = (float*) ((uintptr_t) packed_w + extra_bytes);
    }
    weights += nc * kc;
  } while (--g != 0);
}