// Copyright 2020 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.

#include <xnnpack/assembly.h>

# void xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64(
#     size_t mr,                x0
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const uint8_t*restrict a, x3
#     size_t a_stride,          x4
#     const void*restrict w,    x5
#     uint8_t*restrict c,       x6
#     size_t cm_stride,         x7
#     size_t cn_stride,         [sp] -> x14
$if INC:
  #     const float*restrict acc,  [sp + 8] -> x15
  #     const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 16] -> x8
$else:
  #     const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 8] -> x8

# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.

# A pointers
# x3  a0
# x11 a1
# x12 a2
# x4  a3 / a_stride

# C pointers
# x6  c0
# x9  c1
# x10 c2
# x7  c3 / cm_stride

# Vector register usage
# A0   v0
# A1   v1
# A2   v2
# A3   v3
# B   v20 v21 v22 v23
# C   v16
# C   v18
# C   v28
# C   v30
# Clamp v4, v5, v6
# unused A   v7 v8 v9 v10 v11
# unused B   v19

BEGIN_FUNCTION xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64

        $if INC:
          # Load cn_stride, acc
          LDP x14, x15, [sp]
          # Load params pointer
          LDR x8, [sp, 16]
        $else:
          # Load cn_stride, params pointer
          LDP x14, x8, [sp]

        # Load params values
        LD3R {v4.8h, v5.8h, v6.8h}, [x8]

        # Clamp A and C pointers
        CMP x0, 2                // if mr < 2
        ADD x11, x3, x4          // a1 = a0 + a_stride
        ADD x9, x6, x7           // c1 = c0 + cm_stride
        CSEL x11, x3, x11, LO    //   a1 = a0
        CSEL x9, x6, x9, LO      //   c1 = c0

        ADD x12, x11, x4         // a2 = a1 + a_stride
        ADD x10, x9, x7          // c2 = c1 + cm_stride
                                 // if mr <= 2
        CSEL x12, x11, x12, LS   //   a2 = a1
        CSEL x10, x9, x10, LS    //   c2 = c1

        CMP x0, 4                // if mr < 4
        ADD x4, x12, x4          // a3 = a2 + a_stride
        ADD x7, x10, x7          // c3 = c2 + cm_stride
        CSEL x4, x12, x4, LO     //   a3 = a2
        CSEL x7, x10, x7, LO     //   c3 = c2

0:
        $if INC:
          # Load initial accumulators
          LDP q16, q18, [x15], 32
          LDP q28, q30, [x15], 32
        $else:
          # Load initial bias from w into accumulators
          LDR q16, [x5], 16
          MOV v18.16b, v16.16b
          MOV v28.16b, v16.16b
          MOV v30.16b, v16.16b

        # Is there at least 4 halffloats (8 bytes)?
        SUBS x0, x2, 8  // k = kc - 8
        B.LO 5f

        # Main loop - 4 halffloats of A (8 bytes)
1:
        LDR  d0,  [x3], 8
        LDR q20,  [x5], 16
        LDR q21,  [x5], 16
        LDR  d1, [x11], 8
        LDR  d2, [x12], 8
        LDR  d3,  [x4], 8
        SUBS x0, x0, 8
        FMLA v16.8h, v20.8h, v0.h[0]
        FMLA v18.8h, v20.8h, v1.h[0]
        FMLA v28.8h, v20.8h, v2.h[0]
        FMLA v30.8h, v20.8h, v3.h[0]
        LDR q22, [x5], 16
        LDR q23, [x5], 16

        FMLA v16.8h, v21.8h, v0.h[1]
        FMLA v18.8h, v21.8h, v1.h[1]
        FMLA v28.8h, v21.8h, v2.h[1]
        FMLA v30.8h, v21.8h, v3.h[1]

        FMLA v16.8h, v22.8h, v0.h[2]
        FMLA v18.8h, v22.8h, v1.h[2]
        FMLA v28.8h, v22.8h, v2.h[2]
        FMLA v30.8h, v22.8h, v3.h[2]

        FMLA v16.8h, v23.8h, v0.h[3]
        FMLA v18.8h, v23.8h, v1.h[3]
        FMLA v28.8h, v23.8h, v2.h[3]
        FMLA v30.8h, v23.8h, v3.h[3]
        B.HS 1b

        # Is there a remainder?- 2 halffloats of A (4 bytes)
        TBNZ x0, 2, 6f
        # Is there a remainder?- 1 halffloats of A (2 bytes)
        TBNZ x0, 1, 7f
4:
        # Scale and Clamp
        FMUL v16.8h, v16.8h, v4.8h
        SUBS x1, x1, 8
        FMUL v18.8h, v18.8h, v4.8h
        FMUL v28.8h, v28.8h, v4.8h
        FMUL v30.8h, v30.8h, v4.8h
        FMAX v16.8h, v16.8h, v5.8h
        FMAX v18.8h, v18.8h, v5.8h
        FMAX v28.8h, v28.8h, v5.8h
        FMAX v30.8h, v30.8h, v5.8h
        FMIN v16.8h, v16.8h, v6.8h
        FMIN v18.8h, v18.8h, v6.8h
        FMIN v28.8h, v28.8h, v6.8h
        FMIN v30.8h, v30.8h, v6.8h

        # Store full 4 x 8
        B.LO 8f

        $if INC:
          ST1 {v30.16b},  [x7], x14
          SUB  x3,  x3, x2 // a0 -= kc
          ST1 {v28.16b}, [x10], x14
          SUB x11, x11, x2 // a1 -= kc
          ST1 {v18.16b},  [x9], x14
          SUB x12, x12, x2 // a2 -= kc
          ST1 {v16.16b},  [x6], x14
          SUB  x4,  x4, x2 // a3 -= kc
        $else:
          ST1 {v16.16b},  [x6], x14
          SUB  x3,  x3, x2 // a0 -= kc
          ST1 {v18.16b},  [x9], x14
          SUB x11, x11, x2 // a1 -= kc
          ST1 {v28.16b}, [x10], x14
          SUB x12, x12, x2 // a2 -= kc
          ST1 {v30.16b},  [x7], x14
          SUB  x4,  x4, x2 // a3 -= kc

        B.HI 0b
        RET

5:
        TBZ x0, 2, 7f
6:
        # Remainder- 2 halffloats of A (4 bytes)
        LDR s0,  [x3], 4
        LDR q20, [x5], 16
        LDR q21, [x5], 16
        LDR s1, [x11], 4
        LDR s2, [x12], 4
        LDR s3,  [x4], 4

        FMLA v16.8h, v20.8h, v0.h[0]
        FMLA v18.8h, v20.8h, v1.h[0]
        FMLA v28.8h, v20.8h, v2.h[0]
        FMLA v30.8h, v20.8h, v3.h[0]

        FMLA v16.8h, v21.8h, v0.h[1]
        FMLA v18.8h, v21.8h, v1.h[1]
        FMLA v28.8h, v21.8h, v2.h[1]
        FMLA v30.8h, v21.8h, v3.h[1]

        TBZ x0, 1, 4b

7:
        # Remainder- 1 halffloat of A (2 bytes)
        LDR h0,  [x3], 2
        LDR q20, [x5], 16
        LDR h1, [x11], 2
        LDR h2, [x12], 2
        LDR h3 , [x4], 2
        FMLA v16.8h, v20.8h, v0.h[0]
        FMLA v18.8h, v20.8h, v1.h[0]
        FMLA v28.8h, v20.8h, v2.h[0]
        FMLA v30.8h, v20.8h, v3.h[0]
        B 4b

        # Store odd width
8:
        TBZ x1, 2, 9f
        $if INC:
          STR d30, [x7], 8
          DUP d30, v30.d[1]
          STR d28, [x10], 8
          DUP d28, v28.d[1]
          STR d18, [x9], 8
          DUP d18, v18.d[1]
          STR d16, [x6], 8
          DUP d16, v16.d[1]
        $else:
          STR d16, [x6], 8
          DUP d16, v16.d[1]
          STR d18, [x9], 8
          DUP d18, v18.d[1]
          STR d28, [x10], 8
          DUP d28, v28.d[1]
          STR d30, [x7], 8
          DUP d30, v30.d[1]

9:
        TBZ x1, 1, 10f
        $if INC:
          STR s30,  [x7], 4
          DUP s30, v30.s[1]
          STR s28, [x10], 4
          DUP s28, v28.s[1]
          STR s18,  [x9], 4
          DUP s18, v18.s[1]
          STR s16,  [x6], 4
          DUP s16, v16.s[1]
        $else:
          STR s16,  [x6], 4
          DUP s16, v16.s[1]
          STR s18,  [x9], 4
          DUP s18, v18.s[1]
          STR s28, [x10], 4
          DUP s28, v28.s[1]
          STR s30,  [x7], 4
          DUP s30, v30.s[1]

10:
        TBZ x1, 0, 11f
        $if INC:
          STR h30,  [x7]
          STR h28, [x10]
          STR h18,  [x9]
          STR h16,  [x6]
        $else:
          STR h16,  [x6]
          STR h18,  [x9]
          STR h28, [x10]
          STR h30,  [x7]
11:
        RET

END_FUNCTION xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64

#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif